Retrieval-augmented generation system and method combining metadata filter dual search for financial document retrieval

Abstract

According to the present invention, a search augmentation generation system for financial document search can be provided, comprising: an integrated extraction unit that extracts a plurality of differentiation queries and metadata filters from a user query; a parallel dispatch unit that dispatches vector searches for each of the plurality of differentiation queries asynchronously in parallel; a dual search unit that performs a filter application search and a filter exclusion search, respectively, on the same query embedding vector and merges the results when a domain keyword filter used for partial matching with the document body is included in the metadata filter; and a dual recovery unit that separately recovers the top similarity chunk and the top timeliness chunk from the chunks included in the results of the vector search and removes duplicates. Accordingly, multiple semantic surfaces corresponding to a single query can be recovered simultaneously, and recovery omissions due to the strict application of metadata filters can be compensated for, thereby improving both the recovery rate and precision of financial document search.

Description

Technology Field

[0001] The present invention relates to the field of Retrieval-Augmented Generation (RAG) utilizing Natural Language Processing (NLP) and Large Language Models (LLM).

[0002] More specifically, the present invention relates to a search augmentation generation system and method that simultaneously improves the search recovery rate and precision of financial documents, such as securities firm reports, by extracting together multiple differentiated queries and metadata filters assigned to different semantic axes from a user's natural language query, merging the results by performing application searches and exclusion searches of domain keyword filters in parallel, and separately recovering chunks on the similarity axis and the timeliness axis. Background Technology

[0003] Search augmentation generation is a technology in which, when a user asks a question in natural language, the system searches for relevant documents in a database and uses the retrieved documents as context to have a large-scale language model generate an answer. In this process, the search phase generally operates using two signals: vector similarity search, which converts documents into embedding vectors for storage and retrieves semantically close documents by converting the user query into the same embedding vector; and metadata filtering, which narrows down candidates based on structured attributes attached to the documents, such as stock codes, publication dates, brokerage firm names, and investment opinions.

[0004] Meanwhile, financial documents, such as securities analyst reports, possess essential characteristics that distinguish them from general documents. First, the semantic surfaces regarding the same issue are scattered across multiple axes. For instance, a report addressing a single query like "Samsung Electronics semiconductor business outlook" is written across various semantic surfaces distributed across different dimensions: the keyword dimension, where specific terms appear; the narrative dimension, where content is presented in complete sentences; the industry dimension, representing a sectoral perspective; the individual stock dimension, covering target prices and valuations; and the competitor comparison dimension. When a single query is converted into a single embedding vector, these surfaces are compressed into an average vector; consequently, there was a problem where only chunks from a specific surface were retrieved while the rest were omitted.

[0005] Second, the strict application of metadata filters results in missing retrievals. For example, when a user queries a report regarding a specific product, a general search system retrieves only the chunks where the product name is specified in the metadata. However, the actual report contains numerous chunks where the product's synonym appears in the body text but the product name is not specified in the metadata, or chunks where the product name is missing during the metadata extraction stage. Although these chunks are semantically consistent with the query, they are all excluded from retrieval the moment a strict filter is applied. Conversely, if metadata filters are not applied, many irrelevant chunks are included, leading to a decrease in precision. Consequently, there was a problem where the trade-off between retrieval rate and precision was particularly acute in domain keyword filters.

[0006] Third, the semantic dimension of the financial domain is more granular than that of the general domain. In financial documents, the expression "positive" is essentially separated into the dimension of official investment opinions—specifically the buy, neutral, or sell ratings explicitly assigned by analysts—and the dimension of outlook, which represents the overall tone of the report's body. However, conventional search systems compress this into a single sentiment axis, resulting in the application of filters from a dimension different from the one intended by the user, which in turn retrieves results that differ from the user's intent.

[0007] The following methods are known as conventional technologies that can be applied to such financial document search. First, a single query embedding search method that converts a user query into a single embedding vector for search, and a method (HyDE) that generates a single virtual answer document using a large-scale language model and searches using that embedding are known, but both of these had limitations in that they were limited to retrieving a single semantic surface and did not have a mechanism to generate metadata for search filtering.

[0008] In addition, although multi-query retrieval is known by instructing large-scale language models to generate multiple queries with similar expressions, it was not guaranteed which semantic axis the generated queries belonged to, so many queries often remained variations of expression within a single semantic surface, and it also had the limitation of not being able to produce metadata filters.

[0009] Furthermore, a method known as Self-Query Retrieval involves large-scale language models extracting both search queries and metadata filters from natural language queries to perform pre-filtering prior to similarity searches. However, this method had limitations: first, it produced only a single query without semantic differentiation; second, because it strictly applied extracted metadata filters to pre-compress candidates, chunks that matched the meaning of the text but lacked specified metadata were all omitted from the retrieval process, and there was no mechanism to compensate for this; and third, since it relied on a single sentiment axis, it did not support semantic separation to distinguish between official investment opinions and outlooks.

[0010] Furthermore, although a method for re-ranking search results from multiple queries using Reciprocal Rank Fusion (RRF) scores (RAG-Fusion) is known, it inherited the limitation that queries are not aligned along the semantic axis. Additionally, because it remained a single mechanism of rank-based fusion and lacked an independent retrieval step on the timeliness axis—such as publication date—the most recent documents could be pushed down in similarity rankings and omitted from retrieval, and it had the limitation of being unable to generate metadata filters.

[0011] Therefore, there is a continuing demand for new search augmentation generation technologies that can consistently extract multiple queries with different semantic axes and domain-specific metadata filters from a single natural language query, structurally compensate for missing retrievals resulting from the strict application of metadata filters, and secure both semantic similarity and timeliness. Prior art literature

[0012] Republic of Korea Registered Patent No. 10-3003591 The problem to be solved

[0013] The present invention aims to solve the problems of the aforementioned prior art by generating multiple differentiated queries, each assigned to a different predefined semantic axis from a user's natural language query, and performing vector search on each of them. This overcomes the limitation of a single query embedding recovering only one semantic surface and aims to improve the search recovery rate by simultaneously recovering multiple semantic surfaces corresponding to a single query.

[0014] In addition, the present invention aims to prevent errors caused by the application of filters of unintended dimensions and to improve search precision by determining the modification target of a user expression, deterministically separating and extracting the dimension of official investment opinion and the dimension of outlook or tone, and prohibiting arbitrary inference between filters.

[0015] Furthermore, the present invention aims to improve the recovery rate without loss of precision by compensating for the problem where chunks that match the meaning of the text but do not specify metadata are omitted from the retrieval due to the strict application of metadata filters, by performing filter-applied searches and filter-excluded searches on the same query embedding vector in parallel when a domain keyword filter exists and merging the results.

[0016] In addition, the present invention aims to reduce response delay by asynchronously dispatching vector searches for multiple differentiation queries in parallel, and to secure the latestness of documents along with semantic relevance by separately recovering the top similarity chunk and the top timeliness chunk from the search results. means of solving the problem

[0017] According to one embodiment of the present invention, a search augmentation generation system for financial document search can be provided, comprising: an integrated extraction unit that extracts a plurality of differentiation queries and metadata filters from a user query; a parallel dispatch unit that receives the plurality of differentiation queries and the metadata filters and dispatches a vector search for each of the plurality of differentiation queries asynchronously in parallel; a dual search unit that, when the metadata filter includes a domain keyword filter used for partial matching with the document body, performs a filter application search with the domain keyword filter applied and a filter exclusion search excluding the domain keyword filter, respectively, on the same query embedding vector generated from the differentiation query, and merges the result of the filter application search and the result of the filter exclusion search; and a dual recovery unit that separately recovers the top similarity chunk and the top timeliness chunk from the chunks included in the result of the vector search and removes duplicates.

[0018] Here, the plurality of differentiation queries are generated by being assigned to different predefined semantic axes, and the predefined semantic axes consist of five semantic axes: keyword expansion, virtual answer, industry-centered, stock-centered, and industry or competitor context, and the plurality of differentiation queries may be five differentiation queries, each assigned to one of the five semantic axes.

[0019] In addition, the differentiation query assigned to the semantic axis of the virtual answer is generated as a descriptive sentence by a large-scale language model, but is generated according to constraint prompt rules that prohibit the generation of inference and prediction expressions, and the differentiation query assigned to the semantic axis of the keyword expansion is generated by breaking down the expression of the user query into predefined domain terms without adding new concepts, and if the user query includes a stock code, it can be generated so that the stock code is included in each of the five differentiation queries.

[0020] In addition, the integrated extraction unit may extract the metadata filter according to a formula target determination rule, wherein if the formula target of an expression in a user query is an official investment opinion, an investment opinion filter is set; if the formula target is a forecast or tone, a forecast filter is set; if both an official investment opinion and a forecast or tone are specified in the formula target, both the investment opinion filter and the forecast filter are set; and if the formula target is unclear, neither the investment opinion filter nor the forecast filter is set.

[0021] In addition, the integrated extraction unit may enforce a rule prohibiting inference between filters on the prompt of a large-scale language model so that it is prohibited to infer the value of the forecast filter from the value of the investment opinion filter or to infer the value of the investment opinion filter from the value of the forecast filter.

[0022] In addition, the integrated extraction unit can simultaneously produce the plurality of differentiation queries and the metadata filters within a single response by calling a large-scale language model once with a predefined single structured schema as a constraint.

[0023] In addition, the dual search unit applies the entire metadata filter including the domain keyword filter in the filter-applied search and applies the remaining metadata filter excluding the domain keyword filter in the filter-excluded search, provided that the exact matching filter including the stock code and date range can be applied commonly to both the filter-applied search and the filter-excluded search.

[0024] In addition, the dual search unit may merge the results of the filter-applied search and the filter-excluded search using document identifiers and chunk text as keys, and if the same chunk is retrieved duplicately in both searches, it may prioritize the search result with the smaller vector distance and sort the merged results in order of vector distance.

[0025] In addition, the above dual retrieval unit separately retrieves the chunk with the highest vector distance and the chunk with the latest publication date from the search results of each differentiation query, excludes chunks corresponding to predefined excluded publishers from the retrieval target, selects the next-ranked chunk, and can remove duplicates from the retrieved chunks based on the document identifier.

[0026] According to another embodiment of the present invention, a search augmentation generation method performed in a search augmentation generation system for financial document search can be provided, comprising: an integrated extraction unit extracting a plurality of differentiation queries and metadata filters from a user query; a parallel dispatch unit dispatching a vector search for each of the plurality of differentiation queries asynchronously in parallel; a dual search unit, when the metadata filter includes a domain keyword filter used for partial matching with the document body, performing a filter application search with the domain keyword filter applied and a filter exclusion search excluding the domain keyword filter, respectively, on the same query embedding vector generated from the differentiation query, and merging the result of the filter application search and the result of the filter exclusion search; and a dual recovery unit separately recovering the top similarity chunk and the top timeliness chunk from the chunks included in the result of the vector search and removing duplicates. Effects of the invention

[0027] According to the present invention, since a user's natural language query is forcibly differentiated into multiple differentiated queries, each assigned to a different predefined semantic axis, unlike single query embedding search or multi-query generation methods where semantic axes are not guaranteed, chunks of financial documents distributed across multiple semantic surfaces corresponding to a single query can be retrieved without omission, thereby improving the search retrieval rate.

[0028] In addition, according to the present invention, the dimension of the official investment opinion and the dimension of the outlook or tone are deterministically separated and extracted by determining the modification target of the user expression, and arbitrary inference between filters is prohibited. Therefore, unlike the conventional method of compressing and processing the two dimensions into a single sentiment axis, it is possible to prevent errors caused by arbitrarily filling in filters of dimensions not specified by the user and to improve search precision.

[0029] Furthermore, according to the present invention, when a domain keyword filter exists, filter-applied search and filter-excluded search for the same query embedding vector are performed in parallel, and the results are merged based on vector distance. Therefore, unlike conventional pre-filtering methods that strictly apply extracted metadata filters and omit chunks where the meaning of the text is correct but metadata is not specified, this invention preserves the precise retrieval of chunks with specified metadata while additionally retrieving chunks with missing metadata, thereby simultaneously securing precision and recovery rate. Moreover, since this dual search is performed with the same query embedding on the same vector index, the construction of a separate sparse index is not required.

[0030] In addition, according to the present invention, vector searches for multiple differentiation queries are dispatched asynchronously in parallel so that the total response delay converges to the maximum value of the individual search time, and the top similarity chunk and the top timeliness chunk are retrieved separately from each search result, thereby preventing the problem of the latest documents being pushed down by the similarity ranking in the conventional method that remained in a single mechanism of ranking-based fusion, and providing search results that possess both semantic relevance and timeliness.

[0031] In addition, according to the present invention, since a plurality of differentiated queries and metadata filters are simultaneously produced in a single call to a large-scale language model that has a single structured schema as a constraint, semantic inconsistency that may occur when queries and filters are produced in separate calls is prevented, and the number of language model calls is reduced, thereby reducing processing delay and computation costs.

[0032] The effects of the present invention are not limited to those mentioned above, and other unmentioned technical effects will be clearly understood by a person skilled in the art from the description below. Brief explanation of the drawing

[0033] FIG. 1 is a block diagram illustrating the configuration of a search augmentation generation system for financial document search according to one embodiment of the present invention. FIG. 2 is a conceptual diagram illustrating the process of an integrated extraction unit according to an embodiment of the present invention simultaneously producing a plurality of differentiation queries and metadata filters as a single response. FIG. 3 is a conceptual diagram illustrating semantic dimension separation extraction by formula target determination according to an embodiment of the present invention. FIG. 4 is a conceptual diagram illustrating metafilter dual search and distance-based merging according to an embodiment of the present invention. FIG. 5 is a conceptual diagram illustrating the operation of a parallel dispatch unit, a dual search unit, and a dual retrieval unit according to an embodiment of the present invention. FIG. 6 is a flowchart illustrating a search augmentation generation method for financial document search according to an embodiment of the present invention. Specific details for implementing the invention

[0034] Embodiments of the present invention are described below with reference to the attached drawings so that those skilled in the art can easily implement the invention. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein.

[0035] The terms used herein are for describing the embodiments and are not intended to limit the invention. In this specification, the singular form includes the plural form unless the context clearly indicates otherwise.

[0036] As used herein, “comprises” and “comprising” do not exclude the presence or addition of one or more other components, steps, actions, and / or elements to the mentioned components, steps, actions, and / or elements.

[0037] Furthermore, the components shown in the embodiments of the present invention are illustrated independently to represent different characteristic functions and do not imply that each component consists of separate hardware or a single software unit. That is, for convenience of explanation, each component is described by listing it as a separate component, and at least two of the components may be combined to form a single component, or a single component may be divided into multiple components to perform a function; such integrated and separated embodiments of each component are also included within the scope of the present invention as long as they do not deviate from the essence of the invention.

[0038] In addition, in describing the present invention, if it is determined that a detailed description of related known technology may obscure the essence of the present invention, such detailed description is omitted.

[0039] FIG. 1 is a block diagram illustrating the configuration of a search augmentation generation system for financial document search according to one embodiment of the present invention.

[0040] Referring to FIG. 1, a search augmentation generation system (100) according to one embodiment of the present invention may be configured to include an integrated extraction unit (110), a filter parsing unit (120), a parallel dispatch unit (130), a dual search unit (140), and a dual recovery unit (150).

[0041] The search augmentation generation system (100) can perform a search pipeline in which financial documents, such as securities firm analyst reports, are divided into chunks and each chunk is converted into an embedding vector and stored in a vector database, and the system receives a natural language query from a user, retrieves chunks that match the query, and delivers the retrieved chunks to the context of the answer generation stage. At this time, metadata such as stock code, company name, securities firm, analyst, outlook or tone, target price change, official investment opinion, product or service, and issue date may be attached and stored in each chunk.

[0042] The integrated extraction unit (110) can extract multiple differentiated queries and metadata filters from a user query. Here, the user query may be composed including the user's original query and detailed investigation items, and the integrated extraction unit (110) can configure it in a form to be injected into a prompt of a large-scale language model, and then call the large-scale language model to simultaneously produce differentiated queries and metadata filters.

[0043] The filter parsing unit (120) can convert the metadata filter produced by the integrated extraction unit (110) into a structured schema that the search system can recognize. For example, schema verification can be performed to convert the value produced as a string into a list form or to fill missing fields with null values, thereby allowing the subsequent parallel dispatch unit (130) and dual search unit (140) to share the same filter object.

[0044] The parallel dispatch unit (130) receives multiple differentiation queries and metadata filters and can dispatch vector searches for each of the multiple differentiation queries asynchronously in parallel. The dual search unit (140) can perform filter application search and filter exclusion search in parallel and merge the results when a domain keyword filter is included in the metadata filter, and the dual retrieval unit (150) can separately retriev the top similarity chunk and the top timeliness chunk from the results of the vector search, remove duplicates, and deliver the final retrieval result to the answer generation stage. The detailed operation of each component will be described later with reference to FIGS. 2 to 5.

[0046] FIG. 2 is a conceptual diagram illustrating the process of an integrated extraction unit according to an embodiment of the present invention simultaneously producing a plurality of differentiation queries and metadata filters as a single response.

[0047] Referring to FIG. 2, the integrated extraction unit (110) receives the original query and detailed investigation items from the user and calls a large-scale language model once with a predefined single structured schema, such as a JSON schema, as a constraint, thereby simultaneously producing multiple differentiated queries and metadata filters within a single response. For example, when the user inputs the original query "Analyze the Samsung Electronics (005930) Mirae Asset report" and the detailed investigation item "Samsung Electronics (005930) semiconductor business outlook and performance," the single response of the large-scale language model may include a query array containing five differentiated queries and a filter object containing fields such as stock code, company name, securities firm, date range, product or service.

[0048] Multiple differentiation queries can be generated by assigning each to a different predefined semantic axis. In one embodiment, the predefined semantic axes may consist of five semantic axes: keyword expansion, virtual answer, industry-centric, stock-centric, and industry or competitor context, and multiple differentiation queries can be generated as five differentiation queries, each assigned one to a separate of the five semantic axes.

[0049] Specifically, the first differentiation query assigned to the semantic axis of keyword expansion is a query that breaks down the abstract expression of a user query into concrete terms within the financial domain; for instance, it can map the abstract term "performance" to terms such as "sales revenue," "operating profit," and "net profit." This domain dictionary mapping enhances the semantic cohesion of embedding vectors, thereby contributing to the recovery of keyword-level chunks where concrete terms appear.

[0050] The second differentiation query assigned to the semantic axis of the virtual answer can be generated by a large-scale language model as a descriptive virtual answer of approximately 3 to 4 sentences that follows the tone of a securities firm report. Since the body of financial documents consists of descriptive sentences rather than a simple list of keywords, using descriptive virtual answers in embeddings can improve semantic consistency with body chunks.

[0051] The third differentiation query assigned to the industry-centered semantic axis is a query from an industry analysis perspective, such as business conditions, industry trends, and sector outlook; the fourth differentiation query assigned to the stock-centered semantic axis is a query from a stock analysis perspective, such as individual stock earnings, target prices, and valuations; and the fifth differentiation query assigned to the industry or competitor context semantic axis may be a query from a competitive landscape analysis perspective, such as related industry trends and competitor comparisons. Since these semantic axes are not differentiated into an arbitrary number but are predefined based on the analysis of the perspective of financial document creation, the diversity of the generated differentiation queries can be semantically guaranteed.

[0052] In this case, constraint rules regarding the generation of differentiated queries can be enforced at the prompts of the large-scale language model. For example, it may be enforced to generate exactly five queries and to assign each query to a predefined semantic axis. Additionally, differentiated queries assigned to the keyword expansion semantic axis may be forced to be generated by breaking down the expression of the user query into predefined domain terms without adding new concepts, while differentiated queries assigned to the virtual answer semantic axis may be forced to be generated according to constraint prompt rules that prohibit the generation of inference and prediction expressions such as "it is judged to be..." or "it is expected to be...", as well as a self-verification checklist immediately prior to the response. Through this, virtual answers that align with the descriptive tone of the financial document body can be generated, rather than speculative sentences that are easily derived by general virtual answer generation methods.

[0053] Additionally, if a user query includes a stock code, the system can be generated so that the stock code is included in each of the five differentiated queries. Since semantic axes centered on industry, industry, or competitor contexts tend to expand their meaning toward the general industry, stock identifiers are prone to being dropped during the query rewriting process. However, if the stock code is maintained in all differentiated queries, stock specificity is preserved in common while the five semantic axes target different semantic surfaces, thereby preventing the retrieval of chunks of irrelevant stocks. Depending on the embodiment, the system may also be configured to include a corresponding company name along with or in place of the stock code.

[0054] Meanwhile, the metadata filter produced by the integrated extraction unit (110) may be composed of multiple types of fields including stock code, company name, securities firm, analyst, outlook or tone, target price change, official investment opinion, product or service, and date range. Among these, the stock code, securities firm, analyst, date range, etc., may function as an exact matching filter that must exactly match the value with the stored metadata, and the product or service field may function as a domain keyword filter used for partial matching with the document body.

[0055] As such, since multiple differentiation queries and metadata filters are produced simultaneously within a single call to a large-scale language model that uses a single structured schema as a constraint, semantic inconsistencies between the two that may occur when differentiation queries and metadata filters are produced in separate calls can be prevented, and the number of calls to the language model can be reduced, thereby reducing processing delay and computational costs.

[0057] FIG. 3 is a conceptual diagram illustrating semantic dimension separation extraction by formula target determination according to an embodiment of the present invention.

[0058] Referring to FIG. 3, the integrated extraction unit (110) can apply a formula target determination rule to first determine the formula target of an expression within a user query when extracting a metadata filter, and to separate and set an investment opinion filter corresponding to an official investment opinion and a forecast filter corresponding to a forecast or tone according to the determined formula target.

[0059] Specifically, if the target of a user expression is an official investment opinion—for example, if the user queries "positive investment opinion"—a value such as "BUY" may be set for the investment opinion filter, and the outlook filter may not be set. Conversely, if the target of the expression is a outlook or tone—for example, if the user queries "positive outlook"—a value such as "Positive" may be set for the outlook filter, and the investment opinion filter may not be set. If both the official investment opinion and the outlook or tone are specified in the target of the expression, both filters may be set; however, if the target of the expression is unclear, such as "considering investment," neither filter may be set.

[0060] Additionally, the integrated extraction unit (110) can enforce a rule prohibiting inference between filters on the prompt of the large-scale language model so that it is prohibited to infer the value of the forecast filter from the value of the investment opinion filter or to infer the value of the investment opinion filter from the value of the forecast filter. For example, if "buy" is extracted from the investment opinion filter, "positive" is not automatically filled into the forecast filter. Furthermore, a checklist can be enforced as a prompt to self-verify whether the corresponding filter value is directly mentioned in the user input or whether it was inferred from the value of another filter immediately before the response.

[0061] When official investment opinions, outlooks, or tone are compressed and processed into a single sentiment axis, there is a problem in that filters of dimensions unintended by the user are arbitrarily filled, resulting in chunks different from the user's intent being retrieved. In contrast, when two dimensions are deterministically separated and extracted by determining the modification target as in the present invention, only filters of the dimensions specified by the user are applied, thereby improving search precision.

[0063] FIG. 4 is a conceptual diagram illustrating metafilter dual search and distance-based merging according to an embodiment of the present invention.

[0064] Referring to FIG. 4, the dual search unit (140) can perform the following dual search for each differentiation query when the metadata filter includes a domain keyword filter used for partial matching with the document body.

[0065] First, the text of the differentiation query can be input into an embedding model to generate a query embedding vector. Subsequently, two vector searches can be performed on the same generated query embedding vector. The filtered search applies a full metadata filter, including a domain keyword filter; for instance, in addition to stock code and date range conditions, it can retrieve only chunks where product or service metadata partially matches a specific domain keyword. The filter-excluded search applies only the remaining metadata filters, excluding the domain keyword filter, and can retrieve chunks based on semantic similarity regardless of whether the product or service metadata matches. In this case, the exact matching filter, including the stock code and date range, is applied to both the filtered search and the filter-excluded search, thereby allowing irrelevant candidates to be removed in advance from both searches.

[0066] Next, the dual search unit (140) merges the results of the filter-applied search and the filter-excluded search using document identifiers and chunk text as keys, but if the same chunk is retrieved in both searches, it prioritizes the search result with the smaller vector distance and sorts the merged results in order of vector distance to return a predetermined number of the top results. For example, as shown in FIG. 4, chunks A and B, in which product metadata is specified, are retrieved in the filter-applied search, and chunks D and E, in which synonymous expressions of domain keywords appear in the text but metadata is not specified or is written differently, are additionally retrieved in the filter-excluded search, and chunk A, which is retrieved in both, is adopted only once as the result with the smaller vector distance, and then the entire result can be sorted in ascending order of vector distance.

[0067] When candidates are pre-compressed by strictly applying metadata filters, there is a problem in that chunks where the corresponding keyword is not specified in the metadata, even though synonymous expressions of the domain keyword appear semantically in the body, are all omitted from the retrieval. In contrast, when filter-applied searches and filter-excluded searches for the same query embedding vector are performed in parallel and merged based on vector distance as in the present invention, precise retrieval of chunks with specified metadata is preserved, while additionally recovering chunks with missing metadata, thereby supplementing the retrieval rate without loss of precision. Furthermore, since this dual search is performed as two searches using the same query embedding vector within a single vector index, it can operate without constructing a separate sparse index such as BM25.

[0068] Meanwhile, if the metadata filter does not include a domain keyword filter, a standard single search applying the entire metadata filter may be performed. Additionally, according to the embodiment, the fields targeted by the domain keyword filter may be extended to other fields requiring partial matching with the document body, in addition to products or services.

[0070] FIG. 5 is a conceptual diagram illustrating the operation of a parallel dispatch unit, a dual search unit, and a dual retrieval unit according to an embodiment of the present invention.

[0071] Referring to FIG. 5, the parallel dispatch unit (130) can dispatch vector searches for each of the multiple branched queries in parallel and simultaneously through an asynchronous event loop. For example, search requests for each of the five branched queries are executed simultaneously in the asynchronous event loop, and each individual search can be delegated to a dual search of the dual search unit (140). Accordingly, the total response delay can converge to the level of the maximum value of the individual search times rather than the sum of the five search times.

[0072] The dual retrieval unit (150) can separately retrieve the chunk with the highest vector distance and the chunk with the latest publication date from the search results of each differentiation query. Since the similarity axis, represented by the vector distance, and the timeliness axis, represented by the publication date, are essentially orthogonal, it is highly likely that the top of the two axes are different chunks, and the most recently published document may be omitted if only the similarity axis is retrieved. For example, as shown in FIG. 5, document A with the highest vector distance and document B with the latest publication date may be retrieved from the results of the first differentiation query, and document A with the highest vector distance and document C with the latest publication date may be retrieved from the results of the second differentiation query.

[0073] At this time, the dual recovery unit (150) may exclude chunks corresponding to predefined excluded publishers, such as publishers excluded from citation by policy, from the recovery target and select a next-ranked chunk. Subsequently, the dual recovery unit (150) may remove duplicates based on document identifiers for the recovered chunks, for example, for 5 differentiation queries, up to 10 recovery candidates (2 per query), and rearrange the results with removed duplicates in ascending order of vector distance to construct the final recovery result. According to the embodiment, the number of chunks recovered from each axis is not limited to one but can be expanded to multiple higher-ranking chunks.

[0074] The final retrieval result constructed in this manner can be delivered as a chunk set possessing semantic diversity, precision, and recency to the context of the answer generation stage, and a large-scale language model can generate an answer to a user query based on this.

[0076] FIG. 6 is a flowchart illustrating a search augmentation generation method for financial document search according to an embodiment of the present invention.

[0077] Referring to FIG. 6, a user query including the original user query and detailed investigation items can first be entered (S610).

[0078] Next, the integrated extraction unit (110) can extract multiple differentiation queries and metadata filters from a user query using a large-scale language model (S620). At this time, the multiple differentiation queries and metadata filters can be produced simultaneously with a single language model call that has a predefined single structured schema as a constraint, the multiple differentiation queries can be generated by assigning them to different predefined semantic axes, and the metadata filters can be extracted according to a rule for determining modification targets and a rule for prohibiting inference between filters.

[0079] Next, the filter parsing unit (120) can convert the extracted metadata filter into a structured format (S630).

[0080] Next, the parallel dispatch unit (130) can dispatch vector searches for each of the multiple branch queries asynchronously in parallel (S640).

[0081] Next, for each differentiation query, the dual search unit (140) can perform a filter application search and a filter exclusion search for the same query embedding vector in parallel when a domain keyword filter is included in the metadata filter, and merge the results based on vector distance (S650).

[0082] Next, the dual recovery unit (150) can separately recover the highest similarity chunk and the highest timeliness chunk from each search result, remove duplicates based on document identifiers, and then rearrange them in order of vector distance to form the final recovery result (S660).

[0083] Meanwhile, the embodiments of the present invention are not limited to financial document search, and can be extended to other domains having multi-axial semantic surfaces and fine semantic dimensions, such as medical clinical information search and legal data search, by replacing the definition of semantic axes and the schema of metadata filters to suit the domain.

[0084] Although preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above. It is understood that various modifications can be made by those skilled in the art without departing from the essence of the invention as claimed in the claims, and such modifications should not be understood individually from the technical spirit or perspective of the present invention. Explanation of the symbols

[0086] 100: Search Augmentation Generation System 110: Integrated Extraction Unit 120: Filter parsing section 130: Parallel Dispatch Unit 140: Double search unit 150: Double recovery unit

Claims

Claim 1 A search augmentation generation system for financial document search comprises: an integrated extraction unit that extracts a plurality of differentiation queries and metadata filters from a user query; a parallel dispatch unit that receives the plurality of differentiation queries and the metadata filters and dispatches a vector search for each of the plurality of differentiation queries asynchronously in parallel; a dual search unit that, when the metadata filters include a domain keyword filter used for partial matching with the document body, performs a filter application search applying the domain keyword filter and a filter exclusion search excluding the domain keyword filter, respectively, for the same query embedding vector generated from the differentiation queries, and merges the results of the filter application search and the results of the filter exclusion search; and a dual recovery unit that separately recovers the top similarity chunk and the top timeliness chunk from the chunks included in the results of the vector search and removes duplicates. Claim 2 A search augmentation generation system for financial document search, wherein, in claim 1, the plurality of differentiation queries are generated by being assigned to different predefined semantic axes, and the predefined semantic axes consist of five semantic axes: keyword expansion, virtual answer, industry-centered, stock-centered, and industry or competitor context, and the plurality of differentiation queries are five differentiation queries, each assigned one to the five semantic axes. Claim 3 A search augmentation generation system for financial document search according to paragraph 2, wherein the differentiation query assigned to the semantic axis of the virtual answer is generated as a descriptive sentence by a large-scale language model, but is generated according to constraint prompt rules that prohibit the generation of inference and prediction expressions, and the differentiation query assigned to the semantic axis of the keyword expansion is generated by decomposing the expression of the user query into predefined domain terms without adding new concepts, and if the user query includes a stock code, the stock code is generated to be included in each of the five differentiation queries. Claim 4 A search augmentation generation system for financial document search, wherein the integrated extraction unit extracts the metadata filter according to a modification target determination rule, wherein if the modification target of an expression in a user query is an official investment opinion, an investment opinion filter is set; if the modification target is a forecast or tone, a forecast filter is set; if both the official investment opinion and the forecast or tone are specified in the modification target, both the investment opinion filter and the forecast filter are set; and if the modification target is unclear, neither the investment opinion filter nor the forecast filter is set. Claim 5 A search augmentation generation system for financial document search, wherein, in paragraph 4, the integrated extraction unit enforces a rule prohibiting inference between filters on the prompt of a large-scale language model so as to prohibit inferring the value of the forecast filter from the value of the investment opinion filter or inferring the value of the investment opinion filter from the value of the forecast filter. Claim 6 A search augmentation generation system for financial document search according to claim 1, wherein the integrated extraction unit simultaneously generates the plurality of differentiation queries and the metadata filters within a single response by calling a large-scale language model once with a predefined single structured schema as a constraint. Claim 7 A search augmentation generation system for financial document search according to claim 1, wherein the dual search unit applies the entire metadata filter including the domain keyword filter in the filter-applied search and applies the remaining metadata filter excluding the domain keyword filter in the filter-excluded search, wherein the exact matching filter including the stock code and date range is applied commonly to both the filter-applied search and the filter-excluded search. Claim 8 A search augmentation generation system for financial document search according to claim 7, wherein the dual search unit merges the results of the filter-applied search and the results of the filter-excluded search using document identifiers and chunk text as keys, and if the same chunk is retrieved duplicately in both searches, it prioritizes the search result with the smaller vector distance and sorts the merged results in order of vector distance. Claim 9 A search augmentation generation system for financial document search according to claim 1, wherein the dual retrieval unit separately retrieves the chunk with the highest vector distance and the chunk with the latest issue date from the search results of each differentiation query, excludes chunks corresponding to predefined excluded issuers from the retrieval target and selects the next-ranked chunk, and removes duplicates based on document identifiers for the retrieved chunks. Claim 10 A search augmentation generation method performed in a search augmentation generation system for financial document search, comprising: an integrated extraction unit extracting a plurality of differentiation queries and metadata filters from a user query; a parallel dispatch unit dispatching a vector search for each of the plurality of differentiation queries asynchronously in parallel; a dual search unit, when the metadata filter includes a domain keyword filter used for partial matching with the document body, performing a filter applied search with the domain keyword filter applied and a filter excluded search excluding the domain keyword filter, respectively, on the same query embedding vector generated from the differentiation query, and merging the result of the filter applied search and the result of the filter excluded search; and a dual recovery unit separately recovering the top similarity chunk and the top timeliness chunk from the chunks included in the result of the vector search and removing duplicates.

Images