SPAR: Session-based Pipeline for Adaptive Retrieval on Legacy File Systems

The ability to extract value from historical data is essential for enterprise decision-making. However, much of this information remains inaccessible within large legacy file systems that lack structured organization and semantic indexing, making retrieval and analysis inefficient and error-prone. We introduce SPAR (Session-based Pipeline for Adaptive Retrieval), a conceptual framework that integrates Large Language Models (LLMs) into a Retrieval-Augmented Generation (RAG) architecture specifically designed for legacy enterprise environments. Unlike conventional RAG pipelines, which require costly construction and maintenance of full-scale vector databases that mirror the entire file system, SPAR employs a lightweight two-stage process: a semantic Metadata Index is first created, after which session-specific vector databases are dynamically generated on demand. This design reduces computational overhead while improving transparency, controllability, and relevance in retrieval. We provide a theoretical complexity analysis comparing SPAR with standard LLM-based RAG pipelines, demonstrating its computational advantages. To validate the framework, we apply SPAR to a synthesized enterprise-scale file system containing a large corpus of biomedical literature, showing improvements in both retrieval effectiveness and downstream model accuracy. Finally, we discuss design trade-offs and outline open challenges for deploying SPAR across diverse enterprise settings.

💡 Research Summary

The paper “SPAR: Session-based Pipeline for Adaptive Retrieval on Legacy File Systems” addresses a critical challenge in enterprise data management: efficiently extracting value from vast, unstructured legacy file systems. These systems, often containing decades of heterogeneous data (documents, spreadsheets, reports), lack semantic indexing, making traditional search methods inefficient. While Retrieval-Augmented Generation (RAG) pipelines powered by Large Language Models (LLMs) offer a promising solution, conventional RAG requires building and continuously synchronizing a massive, monolithic vector database that mirrors the entire file system. This process is computationally expensive, hard to scale, and prone to synchronization issues.

To overcome these limitations, the authors propose SPAR, a novel conceptual framework that rethinks the RAG architecture for legacy environments. SPAR’s core innovation is a two-stage, session-based pipeline that replaces the static global vector database with a dynamic, on-demand approach.

In the first stage, SPAR constructs a lightweight Metadata Index. Instead of embedding all file contents into vectors, this index captures high-level, semantically meaningful descriptors of the data. It is structured like a relational database, storing file paths along with enterprise-defined tags (e.g., “Project_Alpha,” “Department_Finance,” “DocumentType_Contract”) that reflect organizational knowledge and domain logic. These tags can be hierarchical, enabling granular filtering. This index is significantly cheaper to build and maintain than a full vector database and can be updated in real-time as files change.

The second stage is the Session-based Adaptive Retrieval process. When a user submits a natural language query, SPAR does not immediately search a giant vector space. Instead, it first parses the query to extract key metadata constraints and keywords. It then uses these constraints to query the Metadata Index, performing a fast, coarse-grained filter to retrieve a relevant subset of files. Only this targeted subset of files is then embedded into vectors to create a small, temporary, session-specific vector database. All subsequent RAG operations (similarity search, context passage to the LLM) are performed within this focused database. Once the user session ends, this temporary database is discarded. This method ensures that the computational heavy lifting of vector search is applied only to data likely to be relevant.

The paper provides a theoretical complexity analysis comparing SPAR to standard RAG pipelines. The analysis demonstrates SPAR’s advantages in terms of lower initial construction cost (O(M) vs. O(N), where M is the number of metadata/tags and N is the total number of files, with M « N), reduced synchronization overhead, and more efficient query handling in large-scale environments.

To validate the framework empirically, the authors applied SPAR to a synthesized enterprise-scale file system containing a large corpus of biomedical literature. The performance was compared against a standard LLM-based RAG pipeline built on a global vector database. Results showed that SPAR achieved higher retrieval effectiveness (improved precision and recall) and better downstream task accuracy (e.g., in question-answering). This improvement is attributed to the metadata filtering step, which effectively narrows the search space and reduces noise before the precise but costly vector similarity search.

Finally, the paper discusses the design trade-offs and open challenges. Key strengths of SPAR include its computational efficiency, scalability, improved controllability for users (via metadata/tag filtering), and easier synchronization with the live file system. The primary challenge and dependency is the quality of the Metadata Index. Its performance hinges on the accuracy and comprehensiveness of the tag assignments. Manually tagging legacy files is labor-intensive, necessitating automated methods using LLMs, which themselves introduce cost and potential error. Other open challenges include optimizing for complex multi-faceted queries, generalizing the framework across diverse enterprise domains with different metadata schemas, and further refining the interaction between the metadata filtering and vector retrieval stages. In conclusion, SPAR presents a pragmatic and efficient alternative to conventional RAG, making LLM-powered knowledge retrieval from legacy systems more feasible and controllable.