RAG assistant for project knowledge

Knowledge exists. Finding it is the work.

A decision may be in a document, implementation details in code, and constraints in an old task. People reconstruct the answer manually; an agent without that context may propose an approach the team has already rejected.

I develop this RAG assistant for my own projects. It indexes documentation, code, and working materials, retrieves relevant passages, and supplies context to people or agents. It integrates with LAF while serving a distinct purpose: access to project knowledge.

Retrieval and answers

• Keyword retrieval finds exact terms, identifiers, and names.
• Semantic retrieval finds related formulations.
• Result fusion and optional reranking reorder the retrieved evidence; the effect on quality needs testing on representative questions.
• The LLM receives the retrieved context, with source references for inspection.

What matters operationally

The index must follow changes to its sources. The system tracks provenance and freshness, supports reindexing, and exposes API and MCP access. Evaluation uses representative questions and distinguishes retrieval quality from answer quality: finding the right passage does not guarantee the correct answer.

Company use cases

Engineering-document search, onboarding, internal instructions, and context for development agents. Adoption begins with sources, access rights, update rules, and evaluation criteria. This working implementation is a foundation for adaptation to your task. We begin with a pilot on agreed sources and a set of your team’s questions.

Knowledge freshness is inspectable

The current index records source-version information and update time. API and MCP processes track the index generation and reload search state after reindexing. That addresses the stale-process problem after source files change.

Retrieval combines BM25, semantic embeddings and reciprocal rank fusion, with an optional reranker. Search, answer generation, indexing and diagnostic methods are available. This is my project implementation, not a claim of ready-made connectors for every enterprise store.

A demonstration using your questions

On approved documents, test why a decision was made, which integration constraints apply and where release instructions live. Inspect retrieved passages and versions before assessing the answer. Include an unanswerable question and a changed document: a useful assistant needs to handle both correctly.

Pilot outputs include a source map, evaluation questions, retrieval results and an integration plan. Enterprise permissions, connectors and operating requirements need validation for your environment.

Discuss search across your knowledge base →