Intelligent retrieval method based on large model and vector enhancement technology

Abstract

The invention provides an intelligent retrieval method based on a large model and a vector enhancement technology, and belongs to the technical field of information retrieval. Based on query intention understanding and self-adaptive segmentation of the large model, a natural language understanding technology is used for analyzing a logic unit to ensure semantic integrity and coherence after segmentation; performing multi-granularity index construction and vectorization processing, capturing semantic information in the text, and encoding the semantic information into a fixed-length vector; and analyzing user portraits and intentions, collecting user interaction history, constructing user portraits, and speculating the user intentions.

Description

Technical Field

[0001] The present invention relates to the field of information retrieval technology, and in particular to an intelligent retrieval method based on large models and vector enhancement technology. Background Art

[0002] Limitations of traditional retrieval technology:

[0003] Keyword matching-based retrieval methods are difficult to handle semantic ambiguity (such as synonyms and polysemy problems)

[0004] Traditional vector space models lack the ability to understand context

[0005] Joint retrieval of multimodal data (text, image, video) lacks a unified representation method

[0006] Defects of existing improvement plan:

[0007] The traditional BERT-based retrieval model has limited parameters and is difficult to handle complex semantic reasoning. Cross-modal retrieval often uses feature splicing, fails to achieve deep semantic alignment, and lacks a feedback mechanism for dynamically adjusting retrieval strategies.

[0008] With the development of informatization, various departments have accumulated a large amount of business data. These data are highly professional and scattered in different business systems. Due to the different system construction times and manufacturers, the data is expressed in various ways. The traditional keyword matching search method can no longer meet business needs. Summary of the Invention

[0009] In order to solve the above technical problems, the present invention provides an intelligent retrieval method based on large models and vector enhancement technology, which solves the problem that each department can realize one-click intelligent retrieval of information within the unit while meeting security requirements, thereby improving efficiency and user experience.

[0010] The technical solution of the present invention is:

[0011] An intelligent retrieval method based on large models and vector enhancement technology, which understands query intent and adaptively segments based on large models, uses natural language understanding technology to analyze logical units to ensure semantic integrity and coherence after segmentation; multi-granularity index construction and vectorization processing capture semantic information in text and encode it into fixed-length vectors; user portrait and intent analysis collects user interaction history, constructs user portraits, and infers user intent.

[0012] Here are the steps:

[0013] Step 1: Text data preprocessing and vectorization;

[0014] Step 2: User portrait construction and maintenance;

[0015] Step 3: Question intent analysis based on historical data;

[0016] Step 4: Vector retrieval and sorting;

[0017] Step 5: Summarize and present the answers;

[0018] Step 6: Real-time feedback and continuous learning.

[0019] Further,

[0020] Step 1 specifically includes

[0021] 1) Data collection: Collect the text data to be processed;

[0022] 2) Adaptive segmentation: Apply intelligent algorithms to identify logical units in the text and use natural language understanding technology to analyze the content of each logical unit, ensuring that each segmented text maintains its semantic integrity and coherence as much as possible;

[0023] 3) Vectorization: Convert each logical unit into a vector representation; use an embedding model to capture the semantic information in the text and encode it into a fixed-length vector;

[0024] 4) Store vectors: Store the generated vectors in an efficient vector database for quick retrieval later.

[0025] Step 2 specifically includes

[0026] 1) Data collection: Automatically collect user interaction history; interaction history includes query records, click behavior, and reading time.

[0027] 2) Profile creation: Based on the collected data, a user profile is constructed, covering the user's interests, preferences, frequently used query areas, and special needs.

[0028] 3) Dynamic update: As new data flows in, user profiles are updated regularly to reflect the latest changes in users’ preferences and needs.

[0029] Step three includes:

[0030] 1) Initial analysis: When receiving a new question from a user, we first use a pre-trained large model to perform a preliminary analysis of the question, extracting keywords and potential intent.

[0031] 2) Intention inference: Inferring the user's potential intentions by combining historical data from the user profile;

[0032] 3) Comprehensive analysis: Combine the results of the initial analysis with inferences based on historical data to more accurately understand the user's true intentions.

[0033] Step 4 specifically includes

[0034] 1) Similarity retrieval: Based on the results of intent analysis, similarity retrieval is performed in the vector database to find the most relevant text entries;

[0035] 2) Result sorting: A comprehensive scoring mechanism is used to sort search results, taking into account the relevance, timeliness and user preferences of the content, to ensure that the information provided not only meets the actual needs of users but is also timely.

[0036] Step 5 specifically includes

[0037] 1) Information integration: The sorted results are fed into a large model to integrate relevant information and generate concise and accurate answers;

[0038] 2) Feedback to users: The generated answers are fed back to users through the user interface, supporting multi-round dialogue mode, allowing users to ask further questions or clarify questions, thereby improving the user experience.

[0039] Step 6, specifically including

[0040] 1) Collect feedback: Collect immediate feedback from users;

[0041] 2) Adjustment strategy: Instantly adjust search strategies and recommended content based on user feedback to ensure that the information provided is always close to the user's actual needs;

[0042] 3) Continuous optimization: Continuous optimization of the system is achieved through continuous learning and adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the workflow of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0045] This paper proposes an innovative intelligent retrieval method that combines a large-scale pre-trained language model with vector enhancement technology. The specific steps are as follows:

[0046] Step 1: Text Data Preprocessing and Vectorization

[0047] 1. Data collection: Collect the text data to be processed from various sources (such as websites, document libraries, etc.).

[0048] 2. Adaptive Segmentation: Intelligent algorithms are applied to identify logical units within the text, such as topic transition points or semantic paragraph boundaries. Natural language understanding techniques (such as semantic role labeling or topic modeling) are used to analyze the content of each logical unit, ensuring that each segmented text maintains its semantic integrity and coherence as much as possible.

[0049] 3. Vectorization: Convert each logical unit into a vector representation. Use an embedding model to capture the semantic information in the text and encode it into a fixed-length vector.

[0050] 4. Store vectors: Store the generated vectors in an efficient vector database for quick retrieval later.

[0051] Step 2: User portrait construction and maintenance

[0052] 1. Data collection: Automatically collect user interaction history, including query records, click behavior, reading time, etc.

[0053] 2. Profile creation: Based on the collected data, a user profile is constructed, covering the user's interests, preferences, frequently used query areas, and special needs.

[0054] 3. Dynamic update: As new data flows in, user profiles are updated regularly to reflect the latest changes in users’ preferences and needs.

[0055] Step 3: Question Intent Analysis Based on Historical Data

[0056] 1. Initial analysis: When receiving a new question from a user, we first use a pre-trained large model to perform a preliminary analysis of the question and extract keywords and potential intent.

[0057] 2. Intent inference: Combine historical data in user portraits to infer users’ potential intentions.

[0058] 3. Comprehensive analysis: Combine the results of the initial analysis with inferences based on historical data to more accurately understand the user's true intent.

[0059] Step 4: Vector retrieval and sorting

[0060] 1. Similarity retrieval: Based on the results of intent analysis, similarity retrieval is performed in the vector database to find the most relevant text entries.

[0061] 2. Result sorting: A comprehensive scoring mechanism is used to sort search results, taking into account factors such as content relevance, timeliness, and user preferences, to ensure that the information provided not only meets the user's actual needs but is also highly timely.

[0062] Step 5: Summarize and present the answers

[0063] 1. Integrate information: Input the sorted results into a large model to integrate relevant information and generate concise and accurate answers.

[0064] 2. Feedback to users: The generated answers are fed back to users through the user interface, supporting multi-round dialogue mode, allowing users to further ask questions or clarify questions, thereby improving the user experience.

[0065] Step 6: Real-time feedback and continuous learning

[0066] 1. Collect feedback: The system can collect real-time feedback from users

[0067] 2. Adjust strategy: Instantly adjust search strategies and recommended content based on user feedback to ensure that the information provided is always close to the user's actual needs.

[0068] 3. Continuous optimization: Through continuous learning and adjustment, the system can be continuously optimized to better meet the personalized needs of users.

[0069] The above description is only a preferred embodiment of the present invention and is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. An intelligent retrieval method based on large models and vector enhancement technology, characterized in that: Based on query intent understanding and adaptive segmentation of large models, natural language understanding technology is used to analyze logical units to ensure semantic integrity and coherence after segmentation; multi-granularity index construction and vectorization processing capture semantic information in the text and encode it into fixed-length vectors; user portrait and intent analysis, collect user interaction history, build user portraits, and infer user intent.

2. The method according to claim 1, characterized in that Here are the steps: Step 1: Text data preprocessing and vectorization; Step 2: User portrait construction and maintenance; Step 3: Question intent analysis based on historical data; Step 4: Vector retrieval and sorting; Step 5: Summarize and present the answers; Step 6: Real-time feedback and continuous learning.

3. The method according to claim 2, characterized in that Step 1 specifically includes 1) Data collection: Collect the text data to be processed; 2) Adaptive segmentation: Apply intelligent algorithms to identify logical units in text and use natural language understanding technology to analyze the content of each logical unit, ensuring that each segmented text maintains its semantic integrity and coherence; 3) Vectorization: Convert each logical unit into a vector representation; use an embedding model to capture the semantic information in the text and encode it into a fixed-length vector; 4) Store vectors: Store the generated vectors in an efficient vector database for quick retrieval later.

4. The method according to claim 2, characterized in that Step 2 specifically includes 1) Data collection: Automatically collect user interaction history; 2) Profile creation: Based on the collected data, a user profile is constructed, covering the user's interests, preferences, frequently used query areas, and special needs; 3) Dynamic update: As new data flows in, user profiles are updated regularly to reflect the latest changes in users’ preferences and needs.

5. The method according to claim 2, characterized in that Step three includes: 1) Initial analysis: When receiving a new question from a user, we first use a pre-trained large model to perform a preliminary analysis of the question, extracting keywords and potential intent. 2) Intention inference: Inferring the user's potential intentions by combining historical data from the user profile; 3) Comprehensive analysis: Combine the results of the initial analysis with inferences based on historical data to more accurately understand the user's true intentions.

6. The method according to claim 2, characterized in that Step 4 specifically includes 1) Similarity retrieval: Based on the results of intent analysis, similarity retrieval is performed in the vector database to find the most relevant text entries; 2) Result sorting: A comprehensive scoring mechanism is used to sort search results, taking into account the relevance, timeliness and user preferences of the content, to ensure that the information provided not only meets the actual needs of users but is also timely.

7. The method according to claim 2, characterized in that Step 5 specifically includes 1) Information integration: The sorted results are fed into a large model to integrate relevant information and generate concise and accurate answers; 2) Feedback to users: The generated answers are fed back to users through the user interface, supporting multi-round dialogue mode, allowing users to ask further questions or clarify questions, thereby improving the user experience.

8. The method according to claim 2, characterized in that Step 6, specifically including 1) Collect feedback: Collect immediate feedback from users; 2) Adjustment strategy: Instantly adjust search strategies and recommended content based on user feedback to ensure that the information provided is always close to the user's actual needs; 3) Continuous optimization: Continuous optimization of the system is achieved through continuous learning and adjustment.

Images