TLDR: The DEREK Module by eSapiens is a secure and scalable Retrieval-Augmented Generation (RAG) pipeline designed for enterprise document question answering. It ingests diverse content, uses a hybrid HNSW+BM25 index, refines queries with GPT-4o, reranks results with Cohere, and generates answers with an LLM, enforcing citation overlap through a LangGraph verifier. This system significantly improves recall and precision, minimizes unsupported statements, and has demonstrated substantial time savings and increased confidence in real-world applications like venture capital due diligence.
In the rapidly evolving landscape of artificial intelligence, businesses are constantly seeking ways to leverage AI for better decision-making and efficiency. A new solution, the DEREK (Deep Extraction & Reasoning Engine for Knowledge) Module, developed by eSapiens, is designed to address a critical need: secure and scalable question answering from vast enterprise document collections.
The DEREK Module is essentially a sophisticated Retrieval-Augmented Generation (RAG) pipeline. RAG systems combine the power of large language models (LLMs) with external knowledge bases to provide more accurate and grounded answers, reducing the common problem of LLM ‘hallucinations’ or making up information.
Addressing Key Enterprise Challenges
eSapiens designed DEREK to tackle several persistent issues faced by organizations:
- Fragmented Knowledge: Enterprise data is often scattered across various formats like PDFs, Office files, web pages, and structured databases. DEREK unifies these disparate sources into a single, queryable knowledge base.
- Recall-Precision Trade-off: Traditional search methods often either miss relevant information (low recall) or return too much irrelevant data (low precision). DEREK employs a hybrid approach, combining vector search (for semantic similarity) with keyword search (for exact matches), to strike a better balance.
- Hallucination and Missing Citations: A major concern with LLMs is their tendency to generate plausible but incorrect information. DEREK includes a unique verification loop that ensures every claim in an answer is directly supported by the retrieved source text, making answers traceable and auditable.
- Operational Overhead: Building and maintaining a production-grade RAG system can be complex and costly. DEREK is delivered as containerized services, simplifying deployment and management while ensuring high security standards like end-to-end TLS 1.3 and AES-256 encryption.
How the DEREK Module Works
The DEREK Module operates through a four-stage pipeline:
1. Ingestion & Chunking: It starts by ingesting diverse content—from unstructured documents like PDFs and Word files to structured data from databases. Documents are then broken down into 1,000-token overlapping chunks, which helps preserve context.
2. Indexing: Each text chunk is converted into numerical representations (embeddings) using advanced models like OpenAI’s ‘text-embedding-3-large’. These are then stored in Elasticsearch, which acts as a hybrid index, supporting both vector-based semantic search and traditional keyword search.
3. Retrieval & Reranking: When a user asks a question, GPT-4o refines the query to improve its clarity and expand synonyms. This enhanced query is then used to retrieve a large set of candidate passages from the hybrid index. These candidates are then re-ranked by Cohere’s ‘rerank-english-v3’ model to identify the most relevant snippets.
4. Generation & Verification: Finally, the top-ranked snippets, along with the refined query, are sent to an LLM. The LLM generates a draft answer using a structured prompt engineering template called CO-STAR. Crucially, a LangGraph-powered verifier then checks every sentence in the draft answer to ensure it is directly supported by the retrieved text. If any claim is ungrounded, the answer is regenerated until all statements are verifiable.
Also Read:
- A RAG Chatbot Enhances Regulatory Compliance for Risk and Quality Assurance
- INRAExplorer: Advancing Scientific Knowledge Exploration with Agentic RAG and Knowledge Graphs
Real-World Impact: Venture Capital Due Diligence
The effectiveness of the DEREK Module was demonstrated in a case study with a venture capital firm. VC teams typically spend significant time manually reviewing documents for due diligence. By deploying DEREK, the firm was able to ingest pitch decks, financial statements, and market reports into a unified, queryable knowledge base.
The results were impressive: manual effort for due diligence was significantly reduced, with insight generation accelerating from weeks to hours. The firm reported closing deals 25–40% faster and evaluating up to 10 times more opportunities with the same team size. Confidence in decisions also improved by 20-30% because every insight was backed by a transparent source, minimizing errors.
The DEREK Module represents a robust step forward in enterprise AI, offering a secure, auditable, and context-faithful solution for document question answering. For more details, you can refer to the original research paper: eSapiens’s DEREK Module: Deep Extraction & Reasoning Engine for Knowledge with LLMs.
