Dynamic context-based search system and method

1 . A computer-implemented method, performed by a server-based query processing module comprising at least one processor and a memory, for dynamic search routing and retrieval using feedback-derived user interaction embeddings, the method comprising:

obtaining a search query via a user interface at a remote user device;

generating a query embedding for the query by executing a transformer-based embedding model, the query embedding representing the search query as a numeric vector;

retrieving relevant feedback-derived user interaction embeddings from a vector database based on similarities between the user interaction embeddings and the query embedding, wherein the user interaction embeddings are numeric vectors encoding feedback parameters extracted from user interface logs and session data for user interface interactions with search results of prior queries;

generating a query contextual classification vector for the search query based on aggregating the retrieved user interaction embeddings, organization metadata, and query parameters derived from the query embedding;

dynamically routing the query embedding to an optimized search model for the search query, wherein the optimized search model is selected from a model registry stored in the memory, the model registry comprising a lexical search model, a semantic search model, and a combined lexical-semantic search model, and wherein selecting the optimized search model comprises:

determining a keyword density score based on a computed similarity between the query contextual classification vector and a keyword-based reference vector;

determining a semantic intent score based on a computed similarity between the query contextual classification vector and a semantic reference vector; and

selecting the lexical search model, the semantic search model, or the combined lexical-semantic search model as the optimized search model based on the keyword density score and the semantic intent score;

processing the query embedding using the optimized search model to retrieve candidate results;

dynamically adjusting a ranking of the candidate results based on computing similarities between the candidate result embeddings and the user interaction embeddings to prioritize candidate results aligned with prior user preferences encoded in the user interaction embeddings; and

displaying, via the user interface at the remote device, the dynamically adjusted candidate results as search results for the query.

2 . The method of claim 1 , wherein selecting the lexical search model comprises:

determining, by a dynamic routing engine, the keyword density score by computing a cosine similarity between the query embedding and the keyword-based reference vector; and

determining the keyword density score is above a keyword density threshold.

3 . The method of claim 1 , wherein selecting the semantic search model comprises:

determining, by a dynamic routing engine, the semantic intent score by computing a cosine similarity between the query contextual classification vector and the semantic reference vector; and

determining the semantic intent score is above a semantic intent threshold.

4 . The method of claim 1 , wherein selecting the combined lexical-semantic search model comprises:

determining, by a dynamic routing engine, the keyword density score by computing a first cosine similarity between the query embedding and the keyword-based reference vector;

determining the keyword density score is above a keyword density threshold;

determining, by a dynamic routing engine, the semantic intent score by computing a second cosine similarity between the query contextual classification vector and the semantic reference vector; and

determining the semantic intent score is above a semantic intent threshold.

5 . The method of claim 1 , further comprising:

monitoring, by the user interface, post-query user interactions to generate an additional interface log and additional session data for the search query;

extracting, by a feedback loop module, additional feedback parameters from the additional interface log and additional session data; and

updating, by the feedback loop module, the vector database to include additional user interaction embeddings generated based on the additional feedback parameters.

6 . The method of claim 5 , wherein extracting the additional feedback parameters from the additional interface log and additional session data comprises extracting, by the feedback loop module, numerical features including one or more of a click-through rate, a dwell time, a bounce rate, and result selection data.

7 . The method of claim 6 , further comprising:

normalizing, by the feedback loop module, the numerical features; and

encoding, by the feedback loop module, the normalized numerical features as the additional user interaction embeddings using a neural network.

8 . The method of claim 5 , wherein updating the vector database comprises:

appending, by a feedback loop module, the additional user interaction embeddings to an append-only memory segment of the vector database and updating a locality-sensitive hash (LSH) table using an incremental indexing operation.

9 . The method of claim 5 , wherein the vector database, managed by the feedback loop module, is user-specific, associating feedback-derived user interaction embeddings with a single user or class of users.

10 . The method of claim 5 , wherein updating the vector database comprises:

adjusting, by a feedback loop module, the user interaction embeddings based on a negative feedback indicator by recomputing cosine similarities to reduce weights of embeddings associated with non-selected results, the negative feedback indicator comprising at least one of a cart abandonment or absence of product selection within a dwell time threshold.

11 . The method of claim 10 , wherein dynamically adjusting the ranking of the candidate results comprises:

increasing, by an embedding-augmented re-ranking model, a candidate rank score of a candidate result associated with a user interaction embedding indicative of a prior product selection.

12 . The method of claim 11 , wherein the embedding-augmented re-ranking model increases the candidate rank score based on determining that a cosine similarity between the candidate result embedding and the user interaction embedding indicative of the prior product selection exceeds a relevance threshold.

13 . The method of claim 5 , further comprising:

prompting, by a feedback loop module via the user interface, the collection of explicit user feedback on search result relevance after a dwell time threshold; and

storing, by the feedback loop module, the explicit feedback as a user interaction embedding in the vector database.

14 . The method of claim 1 , wherein generating the query embedding comprises:

preprocessing, by a query classification module, the search query to produce a normalized query for input to the transformer-based embedding model, the preprocessing including tokenizing the search query using rule-based logic and performing semantic enrichment using a large language model (LLM) executed on a graphics processing unit (GPU) or tensor accelerator.

15 . The method of claim 1 , further comprising:

executing, by a query classification module, a retrieval-augmented classification using a retrieval-augmented generation (RAG)-based model to retrieve the relevant user interaction embeddings and the organizational metadata from the vector database; and

applying, by the query classification module, rule-based weighting to compute a cosine similarity between the query embedding and the retrieved user interaction embeddings.

16 . The method of claim 1 , wherein processing the query embedding using the optimized search model to generate the candidate results comprises executing the selected search model to identify candidate results from one or more product databases.

17 . The method of claim 1 , wherein the user interface logs and session data comprise one or more of result selection data, query refinement data, clickstream logs, bounce rate data, product return logs, dwell time data, scroll depth data, pointer trajectory sampling, and click sequence timing data.

18 . The method of claim 1 , wherein determining the query contextual classification vector comprises:

personalizing, by a query classification module, the query contextual classification vector based on hierarchical user data including at least one of user-specific history, department-level history, or enterprise-level history retrieved from the vector database.

19 . The method of claim 1 , wherein generating the query embedding comprises:

enriching, by a semantic analysis module, the query embedding with structured metadata including at least one of product category, product attributes, or organizational rules before selecting the search model.

20 . A dynamic context-based search system, the system comprising:

a user interface configured to obtain a search query at a remote user device;

a vector database configured to store user interaction embeddings, where the user interaction embeddings are numerical vectors encoding feedback parameters extracted from user interface logs and session data for user interface interactions with search results of prior queries

a query classification module configured to:

generate a query embedding for the query by executing a transformer-based embedding model, the query embedding representing the search query as a numeric vector;

retrieve relevant feedback-derived user interaction embeddings from the vector database based on similarities between the user interaction embeddings and the query embedding; and

generate a query contextual classification vector for the search query based on aggregating the retrieved user interaction embeddings, organization metadata, and query parameters derived from the query embedding;

a model registry configured to store search models, the search models comprising a lexical search model, a semantic search model, and a combined lexical-semantic search model;

a dynamic routing engine configured to:

dynamically route the query embedding to a search model from the model registry optimized for the search query; and

process the query embedding using the optimized search model to retrieve candidate results, wherein, to select the optimized search model from the model registry, the dynamic routing engine is configured to:

determine a keyword density score based on a computed similarity between the query contextual classification vector and a keyword-based reference vector;

determine a semantic intent score based on a computed similarity between the query contextual classification vector and a semantic reference vector; and

select the lexical search model, the semantic search model, or the combined lexical-semantic search model as the optimized search model based on the keyword density score and the semantic intent score; and

a feedback loop module configured to dynamically adjust a ranking of the candidate results based on computing similarities between the candidate result embeddings and the user interaction embeddings to prioritize candidate results aligned with prior user preferences encoded in the user interaction embeddings;

wherein the user interface is configured to display the dynamically adjusted candidate results as search results for the query.

21 . A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause a computing system to:

obtain a search query via a user interface at a remote user device;

generate a query embedding for the query by executing a transformer-based embedding model, the query embedding representing the search query as a numeric vector;

retrieve relevant feedback-derived user interaction embeddings from a vector database based on similarities between the user interaction embeddings and the query embedding, wherein the user interaction embeddings are numeric vectors encoding feedback parameters extracted from user interface logs and session data for user interface interactions with search results of prior queries;

generate a query contextual classification vector for the search query based on aggregating the retrieved user interaction embeddings, organization metadata, and query parameters derived from the query embedding;

dynamically route the query embedding to an optimized search model for the search query, wherein the optimized search model is selected from a model registry stored in the memory, the model registry comprising a lexical search model, a semantic search model, and a combined lexical-semantic search model, and wherein the instructions comprise search model selection instructions that are executable to cause the computing system to:

determine a keyword density score based on a computed similarity between the query contextual classification vector and a keyword-based reference vector;

determine a semantic intent score based on a computed similarity between the query contextual classification vector and a semantic reference vector; and

select the lexical search model, the semantic search model, or the combined lexical-semantic search model as the optimized search model based on the keyword density score and the semantic intent score;

process the query embedding using the optimized search model to retrieve candidate results;

dynamically adjust a ranking of the candidate results based on computing similarities between the candidate result embeddings and the user interaction embeddings to prioritize candidate results aligned with prior user preferences encoded in the user interaction embeddings; and

display, via the user interface at the remote device, the dynamically adjusted candidate results as search results for the query.