A competitive intelligence incremental monitoring method, device, equipment and medium

Abstract

The application provides a competitive intelligence incremental monitoring method, device, equipment and medium, and relates to the technical field of competitive intelligence monitoring and analysis. The method comprises the following steps: obtaining historical baseline content, latest content and atomic reports generated in a preset time window of a monitoring object, performing lightweight processing to form a recent atomic report context, inputting the recent atomic report context into a language model to generate a current atomic report, saving the current atomic report and a corresponding asset snapshot, and establishing a mapping relationship; converting the current atomic report, the historical baseline content, the latest content and candidate source data into a unified attribution input structure; after deduplication screening based on a reference relationship, inputting the unified attribution input structure into the language model again to generate a multi-source attribution report, and establishing a reference relationship between the multi-source attribution report and the source data; and through the unified attribution input structure and the deduplication screening, the number of repeated alarms is reduced, a traceable evidence chain is formed, and the stability and intelligence quality of long-term monitoring are improved.

Description

Technical Field

[0001] This invention relates to the field of competitive intelligence monitoring and analysis technology, and in particular to a method, apparatus, equipment and medium for incremental monitoring of competitive intelligence. Background Technology

[0002] With the rapid development of the internet and social media, businesses are facing increasingly fierce market competition, and dynamic monitoring of competitors has become a crucial means for companies to gain a competitive advantage. Competitive intelligence monitoring systems typically require continuous collection and analysis of target websites, social media keywords, search results, and third-party operational metrics to promptly detect changes in competitors' dynamics.

[0003] Currently, various competitive intelligence monitoring solutions exist in existing technologies. Common competitive intelligence monitoring solutions include the following: 1. Website Change Monitoring Solution: Existing website change monitoring tools typically use scheduled tasks to crawl web pages of the target website, obtain the HTML code, body text, or screenshots of the web pages, and compare the current crawl results with the previous crawl results to discover changes to the website.

[0004] 2. Social Media Monitoring Solution: Existing social media monitoring tools collect posts mentioning the target brand on social media platforms by periodically searching for keywords, brand terms, or account activity, and issue alerts or summaries for new posts.

[0005] 3. Data Metric Monitoring Solution: Existing SEO (Search Engine Optimization), traffic, or brand monitoring tools will periodically call third-party data interfaces to obtain and save various operational metric data returned this time, in order to monitor the changing trends of the target object in search ranking, website traffic, user visits, etc.

[0006] 4. Report Generation and Summary Solution: Existing reporting tools or artificial intelligence models can summarize and aggregate the differences in single data collection results, a batch of social media posts, or a section of webpage content to generate a one-time analysis report for users to review.

[0007] 5. Comprehensive Display Solution: Some enterprise intelligence systems display multiple types of data, such as website change information, brand keyword monitoring results, social media discussion data, asset snapshots, and site intelligence, on the same display interface, making it convenient for users to understand the dynamics of competitors from multiple dimensions.

[0008] However, the existing solutions mentioned above still have the following technical shortcomings in the context of continuous competitor monitoring: First, it can only perform mechanical comparisons between two adjacent data collections. Existing webpage diffing technology typically only compares "the content collected this time" and "the content collected last time," failing to make a comprehensive judgment based on changes that have occurred and been reported over a longer period. When a change persists across multiple collection periods, existing solutions lack the ability to remember cross-period changes, making it difficult to accurately identify whether a change constitutes truly new intelligence, resulting in insufficient accuracy in monitoring results.

[0009] Second, the same change is prone to triggering repeated alerts across different periods. Text adjustments, page structure changes, or user discussions on social media on a target website may persist across multiple data collection periods. Existing solutions lack mechanisms for marking and processing reported content, and lack effective means to trace and identify historical changes that have already been analyzed. This easily leads to the repeated generation of similar analysis reports for the same change in subsequent data collection periods, resulting in redundant alert information and increasing the information filtering burden on users.

[0010] Third, there is a lack of traceability between individual analysis reports and aggregated summaries. Existing summarization tools typically only output the aggregated analysis conclusions without saving details about which individual change reports, which data collection snapshots, or which data sources were specifically cited. When users need to verify the summary conclusions, they cannot trace back to the original evidence from the summary, resulting in insufficient credibility and auditability of the analysis results.

[0011] Fourth, multi-source data lacks a unified attribution input structure. Data structures for various data types, such as website difference analysis results, brand keyword monitoring results, social media discussion content, asset snapshots, and site intelligence, differ significantly. Existing solutions typically only combine and display these data at the presentation layer, lacking a mechanism to uniformly convert these heterogeneous data into a standardized input structure that can be directly used in subsequent analysis processes (such as deduplication, intelligent summarization, and citation linking). This results in low efficiency in attribution analysis of multi-source data.

[0012] Fifth, low-value data lacks a closed-loop processing status. In actual monitoring, the collected data often contains a large number of changes or discussions that have not yielded valuable analytical conclusions. When a batch of data fails to generate effective insights, existing solutions typically do not record the "analyzed but without effective conclusions" status, causing subsequent monitoring tasks to repeatedly process the same batch of low-value data, resulting in a waste of computing resources.

[0013] Sixth, insufficient utilization of long-term monitoring context. Existing solutions mostly generate alarms or summaries on a per-task basis, lacking a mechanism to input historical snapshots, recently generated atomic reports, and current acquisition results into the analysis process. This lack of contextual memory means that each analysis task is performed in isolation, failing to fully utilize historical information accumulated over long-term monitoring to assist in current judgments, thus affecting the stability and accuracy of continuous monitoring tasks and analysis.

[0014] In summary, existing competitive intelligence monitoring technologies have significant shortcomings in areas such as cross-cycle change identification, recurring alarm suppression, multi-source data attribution, evidence chain tracing, and closed-loop analysis. There is an urgent need for a technology solution that can address these issues. Summary of the Invention

[0015] The technical problem to be solved by this invention is to provide a method, device, equipment and medium for incremental monitoring of competitive intelligence. By combining historical baselines with the context of recent atomic reports, it can achieve cross-cycle duplicate identification. By unifying the attribution input structure and deduplication screening, it can reduce duplicate alarms and form a traceable evidence chain, thereby improving the stability and intelligence quality of long-term monitoring.

[0016] In a first aspect, the present invention provides a method for incremental monitoring of competitive intelligence, comprising: Obtain the historical baseline content of the monitored object, wherein the historical baseline content is the most recently collected and saved asset snapshot; Retrieve the latest content of the monitored object in the current period; Obtain the atomic reports generated by the monitored object within a preset time window, and perform lightweight processing on the atomic reports to form a recent atomic report context; The historical baseline content, the latest content, and the recent atomic report context are input into the language model. The language model performs semantic analysis on the input data according to preset structured output constraints and generates the current atomic report. The current atomic report includes at least the set of source identifiers cited in this report. Save the current atomic report and the current asset snapshot corresponding to the latest content, and establish a mapping relationship between the current atomic report and the current asset snapshot; The current atomic report, the historical baseline content, and the latest content, as well as the candidate source data obtained by querying according to the monitoring object, report type, and time range, are respectively converted into a unified attribution input structure. The unified attribution input structure includes at least a source data identifier and a source key. Perform a deduplication filter based on reference relationships on the unified attribution input structure to exclude source data that has been referenced by the same attribution report type, and obtain the deduplicated attribution input; The deduplicated attribution input is sent to the language model. The language model performs multi-source information fusion analysis on the deduplicated attribution input according to the structured output constraints, and generates a multi-source attribution report. The multi-source attribution report includes at least attribution report items and the set of source identifiers they reference. Save the multi-source attribution report and establish a reference relationship between the multi-source attribution report and the multiple source data based on the source identifier set.

[0017] Secondly, the present invention provides a competitive intelligence incremental monitoring device, comprising: The historical baseline recovery module obtains the historical baseline content of the monitored object, which is the most recently collected and saved asset snapshot; The latest content acquisition module retrieves the latest content of the monitored object in the current period; The recent context acquisition module acquires the atomic reports generated by the monitored object within a preset time window, and performs lightweight processing on the atomic reports to form a recent atomic report context; The atomic report generation module inputs the historical baseline content, the latest content, and the recent atomic report context into a language model. The language model performs semantic analysis on the input data according to preset structured output constraints and generates the current atomic report. The current atomic report includes at least the set of source identifiers cited in this report. The storage module saves the current atomic report and the current asset snapshot corresponding to the latest content, and establishes a mapping relationship between the current atomic report and the current asset snapshot; The attribution input conversion module converts the current atomic report, the historical baseline content, and the latest content, as well as the candidate source data obtained by querying according to the monitoring object, report type, and time range, into a unified attribution input structure. The unified attribution input structure includes at least a source data identifier and a source key. The deduplication filtering module performs deduplication filtering based on reference relationships on the unified attribution input structure, excluding source data that has been referenced by the same attribution report type, and obtains the deduplicated attribution input; The multi-source attribution generation module sends the deduplicated attribution input to the language model. The language model performs multi-source information fusion analysis on the deduplicated attribution input according to the structured output constraints and generates a multi-source attribution report. The multi-source attribution report includes at least attribution report items and the set of source identifiers they reference. The reference relationship establishment module saves the multi-source attribution report and establishes reference relationships between the multi-source attribution report and the multiple source data based on the source identifier set.

[0018] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect.

[0019] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0020] One or more technical solutions provided by this invention have at least the following technical effects or advantages: 1. Achieve cross-cycle change identification and improve monitoring accuracy. This invention acquires the historical baseline content of the monitored object, the latest content of the current cycle, and the atomic reports generated within a preset time window. These three data points are then input into a language model for analysis. This allows the analysis process to not only identify the differences between the current cycle and the previous cycle but also to reference recently reported changes, thereby accurately determining whether a change constitutes truly new information. This avoids the misjudgments caused by existing solutions that rely solely on mechanical comparisons of adjacent data collection results.

[0021] 2. Effectively suppresses duplicate alarms and reduces information redundancy. This invention employs a deduplication mechanism based on reference relationships. Before generating multi-source attribution reports, it queries the reference relationship repository for the set of source keys referenced by multi-source attribution reports of the same type that are not discarded. This excludes source data already referenced by similar attribution reports, ensuring that data from the same source does not repeatedly enter the same type of attribution analysis task. This effectively avoids the problem of the same change being repeatedly alarmed in different collection cycles.

[0022] 3. Forming a complete and traceable chain of evidence enhances the credibility of analysis results. This invention establishes a mapping relationship between atomic reports and current asset snapshots during the atomic report generation stage, and establishes a referencing relationship between multi-source attribution reports and data from various sources during the multi-source attribution report generation stage. This allows for tracing from multi-source attribution reports back to the atomic reports they are based on, and from atomic reports back to the asset snapshots corresponding to the collected content. This achieves bidirectional traceability between analytical conclusions and original evidence, facilitating user verification and auditing.

[0023] 4. Unify the heterogeneous data structure from multiple sources to reduce the adaptation cost of attribution analysis. This invention unifies various sources of data, such as website difference atomic reports, brand keyword atomic reports, community analysis reports, company fundamental reports, asset snapshots, and site intelligence data, into an attribution input structure that includes source data identifiers and source keys. This allows data of different types and structures to enter the subsequent deduplication, attribution analysis, and reference link building processes in a unified format, avoiding the tedious work of developing adaptation logic separately for each data type.

[0024] 5. Achieve closed-loop processing of low-value data and avoid waste of computing resources. This invention uses a discarded attribution record mechanism. When a batch of source data fails to form a valid analysis conclusion, a discarded attribution record is generated that is not visible to the user. A reference relationship is established with the batch of source data, marking it as "scanned but no valid analysis conclusion has been formed." This allows subsequent monitoring tasks to identify and skip the marked source data, avoiding the repeated processing of low-value data and saving system computing resources.

[0025] 6. Improve the stability and context continuity of continuous monitoring. Even when the deterministic verification of the atomic report ultimately fails, this invention still saves the current asset snapshot to update the historical baseline content, ensuring that the baseline continuity of subsequent monitoring cycles is not affected by a single analysis failure. Simultaneously, by recording the processing status of all source data through a reference relationship repository, this invention enables long-running monitoring tasks to maintain consistency in processing status across cycles, avoiding the unavailability of subsequent tasks due to a single task failure.

[0026] 7. Ensuring the reliability of analytical output and the authenticity of data sources. This invention employs a dual guarantee mechanism of deterministic verification and source identifier verification. On the one hand, it verifies the atomic reports for field integrity, report type, and no-change markers to ensure that the output format meets expectations. On the other hand, it performs mandatory verification on the source identifiers cited in the atomic reports and multi-source attribution reports to ensure that the source identifiers claimed in the reports correspond to the actual source data input in this case. This prevents reports that cite data from non-existent sources from being entered into the database, thus preventing the risk of "fabricating" sources for analytical results from a fundamental mechanism.

[0027] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the device in Embodiment 2 of the present invention. Detailed Implementation

[0030] The overall concept of the technical solution in this application embodiment is as follows: This scheme uses historical asset snapshots as the baseline anchor for continuous monitoring, recently generated atomic reports as contextual memory for cross-cycle judgment, deduplication based on reference relationships as a filtering mechanism to avoid duplicate analysis, a unified attribution input structure as a standardized bridge for multi-source heterogeneous data, and the mandatory establishment and verification of reference relationships as a guarantee for the construction of the evidence chain.

[0031] Specifically, the steps include the following: Step 1: Create monitoring objects and monitoring tasks The system creates monitoring objects for users, including products, competitors, websites, keywords, and social media keywords. Each monitoring object is associated with at least one monitoring task type, including website content collection tasks, social media collection tasks, company fundamental analysis tasks, and brand keyword analysis tasks.

[0032] Each task includes a task type, product identifier, site identifier, user identifier, scheduling parameters, task context, and optional historical snapshot identifier.

[0033] Step 2: Obtain historical snapshots as incremental comparison baselines When performing website monitoring tasks, the system retrieves the most recent asset snapshot as historical input based on the site identifier. Historical content can be restored from a database or object storage, or it can be passed in by the caller as historical content parameters (such as the previous period's Markdown text, webpage structure information, screenshots, social media data collection results, or other structured operational content). The system uses this as the historical baseline input for this monitoring session.

[0034] Key code example (the historical baseline can be passed in as a parameter or restored from object storage, and drives this difference comparison): def resolve_baseline(parameters, last_snapshot_s3_url): "Prioritizes using historical content passed in by the caller; by default, restores the most recent snapshot from object storage." if parameters.get("last_markdown"): return parameters["last_markdown"], parameters.get("last_snapshot_id") if last_snapshot_s3_url: snapshot = storage.get_content(last_snapshot_s3_url) return snapshot.content, snapshot.snapshot_id return "", None baseline_content, baseline_id = resolve_baseline(parameters, last_snapshot_s3_url) diff_result = None if baseline_content.strip() and current_markdown.strip(): diff_result = MarkdownDiff().compare_markdown( old_markdown=baseline_content, new_markdown=current_markdown, baseline_id=baseline_id, ) Step 3: Extract recent atomic reports as context The system retrieves recent atomic reports (such as difference reports or community analysis reports within the last 7 days) generated by the product or competitors according to a preset time window, and performs lightweight processing, retaining only fields such as title, changed content, impact description, topic, and summary to form recent context input.

[0035] This recent context is used to support the cross-cycle repetition identification in step 4; class DiffAnalysisThinReport(BaseModel): The recent difference report has a lightweight structure, retaining only the core fields used for deduplication and comparison. title: str = "" # Title what_changed: List[str] = Field(default_factory=list) # Changed content why_it_matters: List[str] = Field(default_factory=list) # Impact explanation # Convert recent atomic reports into lightweight context for use in step 4 for cross-cycle recurrence identification. range_history_reports = [ DiffAnalysisThinReport.model_validate(r).model_dump() for r in parameters.get("range_reports", []) ] Step 4: Perform current period data collection and incremental analysis The system calls the acquisition engine to obtain the latest content of the currently monitored object, and then converts the current content, historical baseline, and recent context into a structured input object before inputting it into the analysis engine. The structured input object includes at least a data identifier, data type, source identifier, timestamp, text summary, change field, historical baseline identifier, and processed report identifier.

[0036] The analysis engine generates this atomic report according to the preset structured output constraints. The atomic report may include report type, structured analysis data, alarm level, alarm type, sentiment category, business category, whether there is no change, operation content associated with the collected content, and the set of source identifiers cited in this report.

[0037] The system performs deterministic checks on the analysis results (including checks on field integrity, source identifier, report type, no-change flag, and referenced source set). If the checks fail, the system can retry, downgrade to a result indicating no valid change, or record an error status. If the analysis engine determines that there is no valid change, it generates a result marked with no difference, and the system can still save the corresponding snapshot.

[0038] diff_details = getattr(diff_result, "details", []) if diff_resultelse [] if baseline_content and not diff_details: return AnalysisOutput( metadata={"no_diff": True, "reason": "No changes detectedbetween current and previous content"}, reporters=[], ) result = llm.with_structured_output(DiffAnalysisResponseShape).invoke( input=prompt_template.format_messages(**invoke_input) ) For matter in result. Matters: matter.alert_level = fix_alert_level(matter.alert_level,matter.score) matter.alert_type = fix_alert_type(matter.alert_type) allowed_source_ids = {sid for sid in (current_source_id, baseline_id)if sid} validate_required_source_ids(result, allowed_source_ids=allowed_source_ids) Step 5: Save the atomic report and asset snapshot The system saves atomic reports to the analysis report table and saves the collected content or operational content as asset snapshots. When the snapshot data is large, the snapshot body is uploaded to object storage and a reference is saved in the database. The system also saves the mapping relationship between atomic reports and asset snapshots, so that the corresponding collected content can be traced from the report later.

[0039] Step 6: Construct a multi-source unified attribution input When the system needs to generate a multi-source attribution report, it queries candidate source data by product, report type, and time range. Candidate source data may include website difference atomic reports, brand keyword atomic reports, community analysis reports, historical change reports, company fundamental reports, asset snapshots, site intelligence data, and already generated multi-source attribution reports.

[0040] The system converts candidate data into a unified attribution input structure based on source type. This structure includes at least the source data identifier, unified data type, native data type, core structured data, alarm level, alarm type, generation time, source table or source module identifier, and a source key that can be used to establish referencing relationships. For site intelligence, traffic, bounce rate, visit duration, ratings, and original provider data can be merged into a unified object; for brand keyword reports, context from the company fundamentals report, such as company positioning, target users, and core strengths, can be injected.

[0041] Key code examples (unifying heterogeneous source data into an attributable and chainable input structure): class UnifiedAttributionInput(BaseModel): "Unified Attribution Input Structure for Multi-Source Heterogeneous Reports" source_id: str # Source data identifier unified_type: str # Unified data type (site_diff / brand_keyword / community / site_intel...) native_type: str # Primitive data type (original type of the source table / module) data: Dict[str, Any] # Core structured data alert_level: str # Alert level alert_type: str # Alert type generated_at: str # Generation time source_module: str # Source table or source module identifier source_key: str # The source key used to establish the reference relationship def to_unified_input(report) -> UnifiedAttributionInput: Map fields by source type to convert them to uniform attribution input. source_module = report.source_module or "analysis_report" return UnifiedAttributionInput( source_id=report.report_id, unified_type=report.report_type, native_type=report.native_type, data=report.structured_json or report.data, alert_level=report.alert_level or AlertLevel.LOW.value, alert_type=report.alert_type or AlertType.OTHER.value, generated_at=report.generated_at.isoformat(), source_module=source_module, source_key=f"{source_module}:{report.report_id}", ) Step 7: Deduplication and filtering of source data before multi-source attribution reporting The system performs deduplication filtering based on reference relationships on the unified attribution input, excluding source data that has been referenced by the same attribution report type, and generates different attribution reports according to the report type. The exclusion rules include at least the following: if a source data has been referenced by a non-discarded attribution report of the same type, then the source data will not be included in the attribution report task of the same type; if a source data is only referenced by discarded attribution records, then a preset strategy will be used to determine whether to re-analyze it; different attribution report types maintain their own sets of processed sources.

[0042] def filter_unprocessed(candidates, report_type, retry_discarded=False): "Exclude source data that has already been cited in similar non-discarded attribution reports." referenced_keys = citation_repo.get_referenced_source_keys( report_type=report_type, exclude_discarded=True ) discarded_only_keys = citation_repo.get_discarded_only_source_keys(report_type=report_type) result = [] for candidate in candidates: if candidate.source_key in referenced_keys: continue # This has already been referenced in a similar report, skip. if candidate.source_key in discarded_only_keys and not retry_discarded: continue result.append(candidate) return result Step 8: Save the multi-source attribution report and source data citation relationships. The system sends the unified attribution input and product context to the analysis engine. The analysis engine returns a multi-source attribution report according to preset structured output constraints (including at least the report type, report item set, source identifier set referenced by each report item, whether the tag is discarded, the reason for discarding, and the confidence level). The system performs field validation and source identifier validation on the attribution report to ensure that all referenced source identifiers correspond to the unified attribution input and that no source identifiers other than the input are allowed to be added.

[0043] If the validation fails, the system can retry the analysis engine; if the constraints are still not met after the retry, a discarded attribution record or an error status record will be generated, and no user-visible report will be created. The system saves multi-source attribution reports to the attribution report table and creates a citation source record for each attribution report (including at least the attribution report identifier, source data identifier, citation type, citation label, relevance score, source data type, and creation time).

[0044] # 1) Source Identifier Verification: The sources cited in the attribution report must come from the unified attribution input this time, and cannot be added out of thin air. source_by_id = {source.source_id: source for source in unified_inputs} for item in attribution_report.items: for sid in item.cited_source_ids: if sid not in source_by_id: raise SourceIdValidationError(f"cited source {sid} not inattribution input") # 2) Save the attribution report and create a source citation record (chain of evidence) for each citation. report_id = attribution_repo.save(attribution_report) for item in attribution_report.items: for sid in item.cited_source_ids: source = source_by_id[sid] citation_repo.create( attribution_report_id=report_id, source_id=sid, source_key=source.source_key, reference_type=item.report_type, # Reference type reference_label=item.label, # Reference label relevance=item.relevance, # Relevance score source_data_type=source.unified_type, ) Step 9: Write a discarded attribution record when there are no valid aggregation results. If the analytics engine does not return a valid attribution report, the system generates a discarded attribution record. This record is not visible to the user, but it still establishes a reference relationship with the source data involved in this analysis, marking "this batch of source data has been scanned but no valid insights have been formed," to avoid subsequent tasks repeatedly processing the same batch of low-value data.

[0045] if not attribution_report or not attribution_report.items: # Generate discarded attribution records (invisible to users), but still maintain a reference relationship with the source data. discard_id = attribution_repo.save_discarded( product_id=product_id, report_type=report_type, reason="no valid insight", source_fingerprint=hash_sources([src.source_key for src inunified_inputs]), ) for src in unified_inputs: citation_repo.create( attribution_report_id=discard_id, source_id=src.source_id, source_key=src.source_key, reference_type=report_type, is_discarded=True, # Marks this as a discarded reference; subsequent tasks will not process this batch of sources again. ) Step 10: Optional output to presentation layer The system can write eligible atomic reports or multi-source attribution reports to the presentation layer (such as a timeline) based on configurable display strategies (such as intelligence speed configuration, alarm level, report category, noise marking, duplicate marking, etc.). The output of this presentation layer does not change the source data reference relationships or deduplication status.

[0046] Key code examples (filtered according to configurable display strategies before being written to the presentation layer): def matches_display_policy(item, policy) -> bool: "Filter by combination of alarm level, report category, noise / repeating tags, etc." if _is_ignore(item): # Ignore items and never display them. return False if policy.alert_levels and item.alert_level not in policy.alert_levels: return False if policy.report_types and item.report_type not in policy.report_types: return False if policy.filter_noise and getattr(item, "is_noise", False): return False if policy.filter_duplicate and getattr(item, "is_duplicate",False): return False return True For item in reports: if matches_display_policy(item, display_policy): timeline.write(item) # Only writes to the presentation layer, without changing the underlying reference relationships and deduplication status. Complete the steps. Example 1

[0047] like Figure 1 As shown, this embodiment provides a method for incremental monitoring of competitive intelligence, characterized by including: Obtain the historical baseline content of the monitored object, wherein the historical baseline content is the most recently collected and saved asset snapshot; Retrieve the latest content of the monitored object in the current period; Obtain the atomic reports generated by the monitored object within a preset time window, and perform lightweight processing on the atomic reports to form a recent atomic report context; The historical baseline content, the latest content, and the recent atomic report context are input into the language model. The language model performs semantic analysis on the input data according to preset structured output constraints and generates the current atomic report. The current atomic report includes at least the set of source identifiers cited in this report. Save the current atomic report and the current asset snapshot corresponding to the latest content, and establish a mapping relationship between the current atomic report and the current asset snapshot; The current atomic report, the historical baseline content, and the latest content, as well as the candidate source data obtained by querying according to the monitoring object, report type, and time range, are respectively converted into a unified attribution input structure. The unified attribution input structure includes at least a source data identifier and a source key. Perform a deduplication filter based on reference relationships on the unified attribution input structure to exclude source data that has been referenced by the same attribution report type, and obtain the deduplicated attribution input; The deduplicated attribution input is sent to the language model. The language model performs multi-source information fusion analysis on the deduplicated attribution input according to the structured output constraints, and generates a multi-source attribution report. The multi-source attribution report includes at least attribution report items and the set of source identifiers they reference. Save the multi-source attribution report and establish a reference relationship between the multi-source attribution report and the multiple source data based on the source identifier set.

[0048] In this embodiment, preferably, before obtaining the historical baseline content of the monitored object, the method further includes establishing the monitored object and the monitoring task, specifically: The monitored object is associated with at least one monitoring task type, which includes at least one of website content collection task, social media collection task, company fundamentals analysis task, and brand keyword analysis task. Each monitoring task includes a task type, a monitoring object identifier, a site identifier, a user identifier, scheduling parameters, a task context, and an optional historical snapshot identifier; The historical baseline content is Markdown formatted text or structured operational content, obtained by reading database records, downloading object storage files, or receiving parameters passed in by the caller.

[0049] In this embodiment, preferably, the lightweighting of the atomic report specifically involves retaining only the title, changed content, impact description, subject, and summary fields of the atomic report; The candidate source data includes multiple sources such as website difference atomic reports, brand keyword atomic reports, community analysis reports, company fundamentals reports, historical change reports, asset snapshots, and site intelligence data.

[0050] In this embodiment, preferably, before inputting the historical baseline content, the latest content, and the recent atomic report context into the language model, the method further includes: Perform a text difference comparison between the historical baseline content and the latest content to generate difference details, which include at least the location of the change, the content before and after the change, and the type of change. If the difference details are empty, a current atom report with a no-change flag is generated; If the difference details are not empty, the difference details, the recent atomic report context, and the context information of the monitored object are assembled into a structured input object, and the structured input object is input into the language model. After generating the current atomic report, the process also includes performing a deterministic verification on the current atomic report: The deterministic verification includes field integrity verification, source identifier verification, report type verification, no change marker verification, and reference source set verification; If the verification fails, the corresponding exception handling will be executed according to the verification failure type: If the field integrity check fails or the report type check fails, the language model is called again, and the check failure information is input as a prompt word in the call to the regenerated language model. The number of retries shall not exceed a preset threshold. If the source identifier verification fails or the referenced source set verification fails, the current atomic report will be marked as invalid and discarded, and the current atomic report will not be saved. If the verification still fails after the number of retries reaches the preset threshold, the analysis result will be downgraded to a result with no valid change and the error status will be recorded. For discarded current atomic reports, the current asset snapshot corresponding to the latest content is still saved to update the historical baseline content of the monitored object, so as to maintain the baseline continuity of subsequent monitoring cycles; the discarded current atomic reports do not enter subsequent multi-source attribution and do not participate in the transformation of the unified attribution input structure as candidate source data; The source identifier verification is used to ensure that the source identifiers referenced in the current atomic report correspond to the identifiers of the historical baseline content or the source identifiers of the latest content input into the language model.

[0051] In this embodiment, preferably, the current atomic report, the historical baseline content, and the latest content, as well as the candidate source data obtained by querying according to the monitored object, report type, and time range, are respectively converted into a unified attribution input structure, specifically including: For site intelligence data, its traffic data, bounce rate data, access duration data, rating data, and original service provider data are merged into a unified data object; For the brand keyword atomic report, obtain the company fundamentals report associated with the brand keyword atomic report, and inject the company positioning information, target user information and core advantage information in the company fundamentals report as context data into the unified attribution input structure; After the language model generates a multi-source attribution report, it also includes performing source identifier verification on the multi-source attribution report, specifically: The source identifier verification is used to ensure that the source identifier referenced by each attribution report item in the multi-source attribution report corresponds to the source data identifier carried by each source data in the deduplicated attribution input. If the source identifier referenced by any attribution report item is not among all the source data identifiers corresponding to the deduplicated attribution input, the verification fails, and the language model is called again or a discarded attribution record is generated. If the language model does not return a valid multi-source attribution report, or if the multi-source attribution report contains a dropout flag, then a dropout attribution record is generated: The discarded attribution records are not included in the user-visible report output; The discarded attribution records establish reference relationships with all source data participating in this analysis and are stored in the reference relationship repository. This is used to mark that the batch of source data has been scanned and no valid analysis conclusions have been formed, so as to avoid subsequent monitoring tasks from repeatedly processing the marked source data.

[0052] In this embodiment, preferably, the deduplication filtering based on reference relationships specifically includes: Query the reference relationship repository to obtain the set of source keys that have been referenced by multi-source attribution reports of the same type and are not in a discarded state, where the same type refers to reports with the same attribution report type; Query the reference relationship repository to obtain the set of source keys that are referenced only by discarded attribution records; For each candidate source data, if the source key of the candidate source data exists in the set of source keys referenced by the multi-source attribution reports of the same type and not in a discarded state, then skip the candidate source data; If the source key of the candidate source data exists only in the set of source keys referenced only by the discarded attribution records, then a decision is made on whether to re-include the candidate source data in the analysis according to the preset retry strategy. Different attribution report types maintain their own independent sets of processed source keys.

[0053] In this embodiment, preferably, it also includes: Obtain a preset display strategy, which includes alarm level filtering conditions, report category filtering conditions, noise tag filtering conditions, and duplicate tag filtering conditions; The current atomic report and the multi-source attribution report are filtered according to the display strategy, and the reports that conform to the display strategy are written into the timeline of the display layer. The writing to the presentation layer does not change the reference relationship between the current atomic report or the multi-source attribution report and the source data, nor does it change the deduplication status.

[0054] Based on the same inventive concept, this application also provides an apparatus corresponding to the method in Embodiment 1, as detailed in Embodiment 2. Example 2

[0055] like Figure 2 As shown, this embodiment provides a competitive intelligence incremental monitoring device, including: The historical baseline recovery module obtains the historical baseline content of the monitored object, which is the most recently collected and saved asset snapshot; The latest content acquisition module retrieves the latest content of the monitored object in the current period; The recent context acquisition module acquires the atomic reports generated by the monitored object within a preset time window, and performs lightweight processing on the atomic reports to form a recent atomic report context; The atomic report generation module inputs the historical baseline content, the latest content, and the recent atomic report context into a language model. The language model performs semantic analysis on the input data according to preset structured output constraints and generates the current atomic report. The current atomic report includes at least the set of source identifiers cited in this report. The storage module saves the current atomic report and the current asset snapshot corresponding to the latest content, and establishes a mapping relationship between the current atomic report and the current asset snapshot; The attribution input conversion module converts the current atomic report, the historical baseline content, and the latest content, as well as the candidate source data obtained by querying according to the monitoring object, report type, and time range, into a unified attribution input structure. The unified attribution input structure includes at least a source data identifier and a source key. The deduplication filtering module performs deduplication filtering based on reference relationships on the unified attribution input structure, excluding source data that has been referenced by the same attribution report type, and obtains the deduplicated attribution input; The multi-source attribution generation module sends the deduplicated attribution input to the language model. The language model performs multi-source information fusion analysis on the deduplicated attribution input according to the structured output constraints and generates a multi-source attribution report. The multi-source attribution report includes at least attribution report items and the set of source identifiers they reference. The reference relationship establishment module saves the multi-source attribution report and establishes reference relationships between the multi-source attribution report and the multiple source data based on the source identifier set.

[0056] In this embodiment, preferably, before obtaining the historical baseline content of the monitored object, the method further includes establishing the monitored object and the monitoring task, specifically: The monitored object is associated with at least one monitoring task type, which includes at least one of website content collection task, social media collection task, company fundamentals analysis task, and brand keyword analysis task. Each monitoring task includes a task type, a monitoring object identifier, a site identifier, a user identifier, scheduling parameters, a task context, and an optional historical snapshot identifier; The historical baseline content is Markdown formatted text or structured operational content, obtained by reading database records, downloading object storage files, or receiving parameters passed in by the caller.

[0057] In this embodiment, preferably, the lightweighting of the atomic report specifically involves retaining only the title, changed content, impact description, subject, and summary fields of the atomic report; The candidate source data includes multiple sources such as website difference atomic reports, brand keyword atomic reports, community analysis reports, company fundamentals reports, historical change reports, asset snapshots, and site intelligence data.

[0058] In this embodiment, preferably, before inputting the historical baseline content, the latest content, and the recent atomic report context into the language model, the method further includes: Perform a text difference comparison between the historical baseline content and the latest content to generate difference details, which include at least the location of the change, the content before and after the change, and the type of change. If the difference details are empty, a current atom report with a no-change flag is generated; If the difference details are not empty, the difference details, the recent atomic report context, and the context information of the monitored object are assembled into a structured input object, and the structured input object is input into the language model. After generating the current atomic report, the process also includes performing a deterministic verification on the current atomic report: The deterministic verification includes field integrity verification, source identifier verification, report type verification, no change marker verification, and reference source set verification; If the verification fails, the corresponding exception handling will be executed according to the verification failure type: If the field integrity check fails or the report type check fails, the language model is called again, and the check failure information is input as a prompt word in the call to the regenerated language model. The number of retries shall not exceed a preset threshold. If the source identifier verification fails or the referenced source set verification fails, the current atomic report will be marked as invalid and discarded, and the current atomic report will not be saved. If the verification still fails after the number of retries reaches the preset threshold, the analysis result will be downgraded to a result with no valid change and the error status will be recorded. For discarded current atomic reports, the current asset snapshot corresponding to the latest content is still saved to update the historical baseline content of the monitored object, so as to maintain the baseline continuity of subsequent monitoring cycles; the discarded current atomic reports do not enter subsequent multi-source attribution and do not participate in the transformation of the unified attribution input structure as candidate source data; The source identifier verification is used to ensure that the source identifiers referenced in the current atomic report correspond to the identifiers of the historical baseline content or the source identifiers of the latest content input into the language model.

[0059] In this embodiment, preferably, the current atomic report, the historical baseline content, and the latest content, as well as the candidate source data obtained by querying according to the monitored object, report type, and time range, are respectively converted into a unified attribution input structure, specifically including: For site intelligence data, its traffic data, bounce rate data, access duration data, rating data, and original service provider data are merged into a unified data object; For the brand keyword atomic report, obtain the company fundamentals report associated with the brand keyword atomic report, and inject the company positioning information, target user information and core advantage information in the company fundamentals report as context data into the unified attribution input structure; After the language model generates a multi-source attribution report, it also includes performing source identifier verification on the multi-source attribution report, specifically: The source identifier verification is used to ensure that the source identifier referenced by each attribution report item in the multi-source attribution report corresponds to the source data identifier carried by each source data in the deduplicated attribution input. If the source identifier referenced by any attribution report item is not among all the source data identifiers corresponding to the deduplicated attribution input, the verification fails, and the language model is called again or a discarded attribution record is generated. If the language model does not return a valid multi-source attribution report, or if the multi-source attribution report contains a dropout flag, then a dropout attribution record is generated: The discarded attribution records are not included in the user-visible report output; The discarded attribution records establish reference relationships with all source data participating in this analysis and are stored in the reference relationship repository. This is used to mark that the batch of source data has been scanned and no valid analysis conclusions have been formed, so as to avoid subsequent monitoring tasks from repeatedly processing the marked source data.

[0060] In this embodiment, preferably, the deduplication filtering based on reference relationships specifically includes: Query the reference relationship repository to obtain the set of source keys that have been referenced by multi-source attribution reports of the same type and are not in a discarded state, where the same type refers to reports with the same attribution report type; Query the reference relationship repository to obtain the set of source keys that are referenced only by discarded attribution records; For each candidate source data, if the source key of the candidate source data exists in the set of source keys referenced by the multi-source attribution reports of the same type and not in a discarded state, then skip the candidate source data; If the source key of the candidate source data exists only in the set of source keys referenced only by the discarded attribution records, then a decision is made on whether to re-include the candidate source data in the analysis according to the preset retry strategy. Different attribution report types maintain their own independent sets of processed source keys.

[0061] In this embodiment, preferably, it also includes: Obtain a preset display strategy, which includes alarm level filtering conditions, report category filtering conditions, noise tag filtering conditions, and duplicate tag filtering conditions; The current atomic report and the multi-source attribution report are filtered according to the display strategy, and the reports that conform to the display strategy are written into the timeline of the display layer. The writing to the presentation layer does not change the reference relationship between the current atomic report or the multi-source attribution report and the source data, nor does it change the deduplication status.

[0062] Since the apparatus described in Embodiment 2 of the present invention is an apparatus used to implement the method of Embodiment 1 of the present invention, those skilled in the art can understand the specific structure and variations of the apparatus based on the method described in Embodiment 1 of the present invention, and therefore will not be described again here. All apparatuses used in the method of Embodiment 1 of the present invention fall within the scope of protection of the present invention.

[0063] Based on the same inventive concept, this application provides an electronic device embodiment corresponding to Embodiment 1, as detailed in Embodiment 3. Example 3

[0064] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can implement any of the implementation methods in Embodiment 1.

[0065] Since the electronic device described in this embodiment is the device used to implement the method in Embodiment 1 of this application, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in Embodiment 1 of this application. Therefore, how the electronic device implements the method in the embodiment of this application will not be described in detail here. Any device used by those skilled in the art to implement the method in the embodiment of this application falls within the scope of protection of this application.

[0066] Based on the same inventive concept, this application provides a storage medium corresponding to Embodiment 1, as detailed in Embodiment 4. Example 4

[0067] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it can implement any of the implementation methods in Embodiment 1. Example 5

[0068] This embodiment provides a competitive intelligence incremental monitoring system based on historical snapshot baselines and multi-source attribution. This system is used to continuously monitor, incrementally analyze, and perform multi-source attribution of website content, social media discussions, brand keyword performance, and site operation metrics of an enterprise and its competitors. This helps enterprises promptly identify dynamic changes in the competitive environment and generate traceable analytical reports.

[0069] I. System Architecture The incremental competitive intelligence monitoring system used in this embodiment includes the following functional modules: historical baseline acquisition module, latest content acquisition module, recent context acquisition module, atomic report generation module, storage module, attribution input conversion module, deduplication filtering module, multi-source attribution generation module, and reference relationship establishment module. Each of these modules is deployed on one or more servers, which are equipped with processors, memory, network interfaces, and input / output interfaces. The modules interact with each other via a system bus or network communication interface.

[0070] The language model employs a large language model based on the Transformer architecture. The system calls the language model's application programming interface and sets structured output parameters, enabling the language model to output structured analysis results in a preset JSON Schema format. The reference relationship repository uses a relational database, while the main content of asset snapshots is stored in object storage services. The database tables only store the snapshot's metadata and the access address to the object storage.

[0071] II. Establishment of Monitoring Tasks The system creates monitoring objects for users. These objects are the target entities that users need to monitor, including the company's own products, competitors' products, target websites, brand keywords, and social media keywords. Each monitoring object is associated with at least one monitoring task type.

[0072] In this embodiment, it is assumed that a user needs to monitor the official website activity, social media discussion volume, and brand keyword performance of competitor A. The system creates the following monitoring tasks for this user: website content collection task (target: competitor A's official website homepage, collection cycle: once every 6 hours), social media collection task (target: posts mentioning "competitor A" on platforms such as Weibo and Twitter, collection cycle: once every 2 hours), brand keyword analysis task (target: search index of "competitor A" in search engines, collection cycle: once every 24 hours), and company fundamentals analysis task (target: competitor A's public financial reports and news announcements, collection cycle: once every 7 days). Each monitoring task is stored in the system as a task record, including fields such as task type, monitoring object identifier, site identifier, user identifier, scheduling parameters, task context, and historical snapshot identifier.

[0073] III. Acquisition of Historical Baseline Content When the system starts a monitoring task, it first retrieves the most recently saved asset snapshot as the historical baseline content.

[0074] Taking website content collection as an example, the system queries the asset snapshot table to obtain the most recently successfully saved snapshot record corresponding to the site identifier. If a historical snapshot exists, the system downloads the main content of the snapshot from the object storage service based on the object storage address in the snapshot record. If the download fails, the system returns an error status to the caller and records an error log. If no historical snapshot exists in the system, or if the caller passes historical content in the task parameters, the system will prioritize using the content passed by the caller as the historical baseline.

[0075] In this embodiment, the system successfully downloaded the Markdown format text of competitor A's official website homepage from 6 hours ago, which was used as the historical baseline content for this analysis. At the same time, the system obtained the identifier of this snapshot for subsequent source identifier verification.

[0076] IV. Access to the Latest Content The system invokes the data acquisition engine to perform data acquisition for the current period according to the scheduling parameters of the monitoring task. The acquisition engine uses a headless browser or HTTP client to send a request to the target URL to obtain the complete HTML source code of the webpage, parses and cleans the HTML content, extracts the page text, and converts the text content into Markdown format.

[0077] In this embodiment, the latest content returned by the data collection engine is the Markdown format text of competitor A's official website homepage at the current moment. Comparing this latest content with the historical baseline content reveals that a new product, "New Product X," has been added to the product list, while promotional activity information has been deleted.

[0078] V. Recent Atomic Report Context Acquisition The system queries the analysis report table for the atomic reports generated for the monitored object according to the preset time window (set to the last 7 days in this embodiment), and performs lightweight processing on the queried atomic reports, extracting only five core fields: title, changed content, impact description, topic and summary. Other fields are discarded in this lightweight processing.

[0079] In this embodiment, the system found 6 historical atomic reports for the site within the past 7 days. After lightweight processing, the system assembled these 6 reports into a recent atomic report context for subsequent cross-period duplicate identification.

[0080] VI. Current Atomic Report Generation The system performs a textual difference comparison between historical baseline content and the latest content, generating difference details. These details include the location of the change, the content before and after the change, and the type of change. In this example, the difference comparison results show that competitor A's homepage has added product information for "New Product X" (change type: addition) and removed the promotional countdown information (change type: deletion).

[0081] The system assembles the difference details, recent atomic report context, and monitoring object context information into a structured input object, which is then input into the language model. The language model performs semantic analysis on this input according to preset structured output constraints to generate the current atomic report. In this embodiment, the language model, after referring to the recent atomic report context, finds that "New Product X" has not been mentioned in the historical reports of the past 7 days, therefore determining that this change is genuinely new information, rather than a duplicate report. The current atomic report includes report type, structured analysis data, alarm level (important), alarm type (product change), set of referenced source identifiers, and a no-change flag (no).

[0082] After the language model is output, the system performs deterministic checks on the atomic report, including field integrity checks, source identifier checks, report type checks, no-change flag checks, and referenced source set checks. In this embodiment, the atomic report passed all checks. For different types of check failures, the system performs corresponding exception handling: if field integrity or report type checks fail, the language model is called again; if source identifier or referenced source set checks fail, the atomic report is marked as invalid and discarded.

[0083] VII. Preservation of Atomic Reports and Asset Snapshots After the atomic report passes deterministic verification, the system saves it to the analysis report table. Simultaneously, the system saves the latest content returned by the acquisition engine as a new asset snapshot, uploads it to the object storage service, and creates a snapshot record in the asset snapshot table. The system further establishes a mapping relationship between atomic reports and asset snapshots, associating the current atomic report, the current asset snapshot, and historical baseline snapshots in the mapping table. This allows subsequent analysis to trace back to the acquisition content upon which the atomic report was based.

[0084] VIII. Preparations for Generating a Multi-Source Attribution Report The system queries candidate source data by monitoring object, report type, and time range. In this embodiment, the candidate source data queried by the system includes: currently generated website difference atomic reports, 6 historical website difference atomic reports from the past 7 days, 2 brand keyword atomic reports, 3 social media community analysis reports, 1 company fundamentals report, and site intelligence data.

[0085] IX. Unified Attribution Input Transformation The system converts the various candidate source data obtained from the query into a unified attribution input structure. This unified structure includes source data identifier, unified data type, original data type, core structured data, alarm level, alarm type, generation time, source module identifier, and source key.

[0086] The transformation strategies for various data sources are as follows: For website difference atomic reports, changes are directly extracted from the structured analysis data of the reports; for brand keyword atomic reports, in addition to extracting search popularity values ​​and ranking change data, related company fundamental reports are queried from the citation relationship repository, and company positioning information, target user information, and core advantage information are injected as contextual data; for community analysis reports, the number of posts, sentiment classification, and discussion popularity trends are extracted; for site intelligence data, traffic, bounce rate, visit duration, ratings, and original service provider data are merged into a unified data object. After transformation, the system obtains 12 unified attribution input structures.

[0087] 10. Deduplication based on reference relationships The system performs deduplication filtering on the unified attribution input based on reference relationships. First, it queries the reference relationship repository to obtain the set of source keys that have been referenced by multi-source attribution reports of the same type and are not discarded, and the set of source keys that have only been referenced by discarded attribution records. For each candidate source data, if its source key exists in the set "referenced by non-discarded attribution reports of the same type," it is skipped; if it only exists in the set "referenced by discarded attribution records," it is re-included in the analysis according to a preset retry strategy. In this embodiment, the system selects 7 from 12 candidate source data to form the deduplicated attribution input.

[0088] XI. Generation and Storage of Multi-Source Attribution Reports The system inputs the seven deduplicated attribution data into the language model. After multi-source information fusion analysis, the language model generates a multi-source attribution report. In this embodiment, the language model concludes that competitor A recently released "New Product X," simultaneously secured a new round of funding, and is experiencing rising social media discussion and search indexes, while its site bounce rate is decreasing, indicating that competitor A is in a phase of accelerated development.

[0089] The multi-source attribution report includes the attribution report type, the set of attribution report items, and the set of source identifiers they reference. The system performs source identifier verification on the multi-source attribution report to ensure that each source identifier referenced by an attribution report item corresponds to the source data entered in this instance. After successful verification, the system saves the multi-source attribution report to the attribution report table and establishes a reference relationship between each attribution report item and each source data. Each reference record includes the attribution report identifier, source data identifier, source key, reference type, reference tag, relevance score, and source data type.

[0090] 12. Generation of Discarded Attribution Records If the language model does not return a valid multi-source attribution report, or if the attribution report contains discard tags, the system generates a discarded attribution record. This record is not output as a user-visible report, but the system establishes a reference relationship for all source data involved in this analysis and marks it as discarded. This mechanism is used to mark source data that has been scanned but has not yielded valid analytical conclusions, avoiding redundant processing in subsequent tasks.

[0091] 13. Presentation Layer Output The system writes eligible atomic reports or multi-source attribution reports into the timeline of the presentation layer based on preset display strategies (including alarm level filtering, report category filtering, noise tag filtering, and duplicate tag filtering). This operation does not change the referencing relationship between the report and the source data or the deduplication status, thus decoupling the core analysis logic from the presentation logic.

[0092] XIV. Stability Guarantee for Long-Term Monitoring After each monitoring task is completed, regardless of whether the atomic report passes deterministic verification, the system saves the latest collected content as an asset snapshot, which is used to update the historical baseline content of the monitored object. This mechanism ensures that a single analysis failure does not cause the baseline to stagnate, maintaining the baseline continuity for subsequent monitoring cycles.

[0093] Meanwhile, the system persistently saves the processing status of all source data through a reference relationship repository, enabling long-running monitoring tasks to maintain a consistent processing status across different cycles.

[0094] 15. A typical application scenario Assume the monitoring task has been running continuously for 30 days. At 10:00 AM on the 30th day, the system triggers a new data collection cycle, retrieving the latest content from competitor A's official website homepage. Using the asset snapshot from day 29 as the historical baseline and seven atomic reports from the past seven days as the recent context, the data collection results show that competitor A's homepage has added a "New Product X" display area. Language model analysis reveals that "New Product X" was discussed by users in the social media reports from day 28, therefore this related information is noted in the atomic reports.

[0095] The atomic reports are saved after verification. The system uniformly converts today's atomic reports, historical baseline content, historical atomic reports, brand keyword reports, community analysis reports, and site intelligence data. After deduplication, the data from these seven sources is input into a language model to generate a multi-source attribution report. The report states: "Competitor A's 'New Product X' has been officially launched, and there has been pre-launch discussion on social media. The search index has increased by 12% month-on-month. We recommend closely monitoring its pricing strategy and marketing activities." Users can view the report through the presentation layer and click on the reference tags to trace back to the atomic reports and asset snapshots, achieving complete evidence chain tracing.

[0096] Thus, this embodiment realizes a closed-loop process from data acquisition, incremental analysis, multi-source attribution to display output, effectively solving technical problems in the prior art such as cross-cycle repeated alarms, difficulty in multi-source data attribution, untraceable evidence, and repeated processing of low-value data.

[0097] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for incremental monitoring of competitive intelligence, characterized in that, include: Obtain the historical baseline content of the monitored object, wherein the historical baseline content is the most recently collected and saved asset snapshot; Retrieve the latest content of the monitored object in the current period; Obtain the atomic reports generated by the monitored object within a preset time window, and perform lightweight processing on the atomic reports to form a recent atomic report context; The historical baseline content, the latest content, and the recent atomic report context are input into the language model. The language model performs semantic analysis on the input data according to preset structured output constraints and generates the current atomic report. The current atomic report includes at least the set of source identifiers cited in this report. Save the current atomic report and the current asset snapshot corresponding to the latest content, and establish a mapping relationship between the current atomic report and the current asset snapshot; The current atomic report, the historical baseline content, and the latest content, as well as the candidate source data obtained by querying according to the monitoring object, report type, and time range, are respectively converted into a unified attribution input structure. The unified attribution input structure includes at least a source data identifier and a source key. Perform a deduplication filter based on reference relationships on the unified attribution input structure to exclude source data that has been referenced by the same attribution report type, and obtain the deduplicated attribution input; The deduplicated attribution input is sent to the language model. The language model performs multi-source information fusion analysis on the deduplicated attribution input according to the structured output constraints, and generates a multi-source attribution report. The multi-source attribution report includes at least attribution report items and the set of source identifiers they reference. Save the multi-source attribution report and establish a reference relationship between the multi-source attribution report and the multiple source data based on the source identifier set.

2. The method according to claim 1, characterized in that, Before obtaining the historical baseline content of the monitored object, the process also includes establishing the monitored object and monitoring task, specifically: The monitored object is associated with at least one monitoring task type, which includes at least one of website content collection task, social media collection task, company fundamentals analysis task, and brand keyword analysis task. Each monitoring task includes a task type, a monitoring object identifier, a site identifier, a user identifier, scheduling parameters, a task context, and an optional historical snapshot identifier; The historical baseline content is Markdown formatted text or structured operational content, obtained by reading database records, downloading object storage files, or receiving parameters passed in by the caller.

3. The method according to claim 1, characterized in that, The lightweighting process for the atomic report specifically involves retaining only the title, changed content, impact description, subject, and summary fields of the atomic report. The candidate source data includes multiple sources such as website difference atomic reports, brand keyword atomic reports, community analysis reports, company fundamentals reports, historical change reports, asset snapshots, and site intelligence data.

4. The method according to claim 1, characterized in that, Before inputting the historical baseline content, the latest content, and the recent atomic report context into the language model, the following steps are also included: Perform a text difference comparison between the historical baseline content and the latest content to generate difference details, which include at least the location of the change, the content before and after the change, and the type of change. If the difference details are empty, a current atom report with a no-change flag is generated; If the difference details are not empty, the difference details, the recent atomic report context, and the context information of the monitored object are assembled into a structured input object, and the structured input object is input into the language model. After generating the current atomic report, the process also includes performing a deterministic verification on the current atomic report: The deterministic verification includes field integrity verification, source identifier verification, report type verification, no change marker verification, and reference source set verification; If the verification fails, the corresponding exception handling will be executed according to the verification failure type: If the field integrity check fails or the report type check fails, the language model is called again, and the check failure information is input as a prompt word in the call to the regenerated language model. The number of retries shall not exceed a preset threshold. If the source identifier verification fails or the referenced source set verification fails, the current atomic report will be marked as invalid and discarded, and the current atomic report will not be saved. If the verification still fails after the number of retries reaches the preset threshold, the analysis result will be downgraded to a result with no valid change and the error status will be recorded. For discarded current atomic reports, the current asset snapshot corresponding to the latest content is still saved to update the historical baseline content of the monitored object, so as to maintain the baseline continuity of subsequent monitoring cycles; the discarded current atomic reports do not enter subsequent multi-source attribution and do not participate in the transformation of the unified attribution input structure as candidate source data; The source identifier verification is used to ensure that the source identifiers referenced in the current atomic report correspond to the identifiers of the historical baseline content or the source identifiers of the latest content input into the language model.

5. The method according to claim 1, characterized in that, The current atomic report, the historical baseline content, and the latest content, as well as the candidate source data obtained by querying according to the monitored object, report type, and time range, are respectively converted into a unified attribution input structure, specifically including: For site intelligence data, its traffic data, bounce rate data, access duration data, rating data, and original service provider data are merged into a unified data object; For the brand keyword atomic report, obtain the company fundamentals report associated with the brand keyword atomic report, and inject the company positioning information, target user information and core advantage information in the company fundamentals report as context data into the unified attribution input structure; After the language model generates a multi-source attribution report, it also includes performing source identifier verification on the multi-source attribution report, specifically: The source identifier verification is used to ensure that the source identifier referenced by each attribution report item in the multi-source attribution report corresponds to the source data identifier carried by each source data in the deduplicated attribution input. If the source identifier referenced by any attribution report item is not among all the source data identifiers corresponding to the deduplicated attribution input, the verification fails, and the language model is called again or a discarded attribution record is generated. If the language model does not return a valid multi-source attribution report, or if the multi-source attribution report contains a dropout flag, then a dropout attribution record is generated: The discarded attribution records are not included in the user-visible report output; The discarded attribution records establish reference relationships with all source data participating in this analysis and are stored in the reference relationship repository. This is used to mark that the batch of source data has been scanned and no valid analysis conclusions have been formed, so as to avoid subsequent monitoring tasks from repeatedly processing the marked source data.

6. The method according to claim 1, characterized in that, The deduplication filtering based on reference relationships specifically includes: Query the reference relationship repository to obtain the set of source keys that have been referenced by multi-source attribution reports of the same type and are not in a discarded state, where the same type refers to reports with the same attribution report type; Query the reference relationship repository to obtain the set of source keys that are referenced only by discarded attribution records; For each candidate source data, if the source key of the candidate source data exists in the set of source keys referenced by the multi-source attribution reports of the same type and not in a discarded state, then skip the candidate source data; If the source key of the candidate source data exists only in the set of source keys referenced only by the discarded attribution records, then a decision is made on whether to re-include the candidate source data in the analysis according to the preset retry strategy. Different attribution report types maintain their own independent sets of processed source keys.

7. The method according to claim 1, characterized in that, Also includes: Obtain a preset display strategy, which includes alarm level filtering conditions, report category filtering conditions, noise tag filtering conditions, and duplicate tag filtering conditions; The current atomic report and the multi-source attribution report are filtered according to the display strategy, and the reports that conform to the display strategy are written into the timeline of the display layer. The writing to the presentation layer does not change the reference relationship between the current atomic report or the multi-source attribution report and the source data, nor does it change the deduplication status.

8. A competitive intelligence incremental monitoring device, characterized in that, include: The historical baseline recovery module obtains the historical baseline content of the monitored object, which is the most recently collected and saved asset snapshot; The latest content acquisition module retrieves the latest content of the monitored object in the current period; The recent context acquisition module acquires the atomic reports generated by the monitored object within a preset time window, and performs lightweight processing on the atomic reports to form a recent atomic report context; The atomic report generation module inputs the historical baseline content, the latest content, and the recent atomic report context into a language model. The language model performs semantic analysis on the input data according to preset structured output constraints and generates the current atomic report. The current atomic report includes at least the set of source identifiers cited in this report. The storage module saves the current atomic report and the current asset snapshot corresponding to the latest content, and establishes a mapping relationship between the current atomic report and the current asset snapshot; The attribution input conversion module converts the current atomic report, the historical baseline content, and the latest content, as well as the candidate source data obtained by querying according to the monitoring object, report type, and time range, into a unified attribution input structure. The unified attribution input structure includes at least a source data identifier and a source key. The deduplication filtering module performs deduplication filtering based on reference relationships on the unified attribution input structure, excluding source data that has been referenced by the same attribution report type, and obtains the deduplicated attribution input; The multi-source attribution generation module sends the deduplicated attribution input to the language model. The language model performs multi-source information fusion analysis on the deduplicated attribution input according to the structured output constraints and generates a multi-source attribution report. The multi-source attribution report includes at least attribution report items and the set of source identifiers they reference. The reference relationship establishment module saves the multi-source attribution report and establishes reference relationships between the multi-source attribution report and the multiple source data based on the source identifier set.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images