Enterprise RAG knowledge platform

Engineers in the industrial and energy sector need answers buried in thousands of pages of OEM manuals and IOMs. A generic chatbot isn't enough: every answer has to cite the exact page, and each business unit may only see its own documents.

• Agentic hybrid retrieval: multi-turn query rewriting, then a tool-using agent (vector search, keyword search, document reads; capped at 4 turns) over pgvector HNSW and Postgres full-text search fused with RRF, with parent/child context expansion.
• Citation grounding: answers cite page-level sources, structural verification drops any citation that doesn't match a retrieved chunk, and everything streams to the client over SSE.
• Ingestion: Docling with OCR, contextual chunk enrichment, SHA-256 dedup, embedding reuse, and leased workers with retries.
• Tenant isolation: 23 permissions across Global, Organization and Unit scopes, a unit filter on every retrieval query, and isolation tests in both services.
• Auth: JWT bearer tokens for every sign-in method, including Microsoft Entra ID single sign-on.
• Guardrails: retrieved text is treated as data, never instructions; answers come only from cited sources, and the system declines when nothing relevant is found. A rule-based intent check turns away off-topic requests before any retrieval runs.
• Defence in depth: unit-scoped authorization in the API, a unit filter in every SQL retrieval path, and an internal RAG service the browser can never reach directly.
• Re-ingest tooling: embedding model and version tracked on every chunk, with unchanged chunks reusing their stored vectors when documents are reprocessed.
• Provider-agnostic adapters for LLMs, embeddings, rerankers and parsers, running on Microsoft Foundry (chat and text-embedding-3-large).

• pgvector over a dedicated vector DB: one database holds metadata, vectors and tenant filters.
• Hybrid search plus RRF over a reranker alone: keyword search catches exact matches that embeddings miss.
• The browser never calls the RAG service directly; the API layer owns auth and persistence.
• PyMuPDF rejected over its AGPL licence.

• Page drift: the parser invented page numbers, corrupting citations. Fixed with nullable page bounds end to end, embeddings preserved on re-ingest, and a backfill script.
• Concurrency crash: parallel vector and keyword searches on one async DB session failed, so they now run sequentially.
• Hedge text breaking answers: fixed with grounded decoding, with the verified final answer treated as canonical.

• Team and tech lead for 4 engineers: wrote the architecture docs and ERD, set PR review rules and CI quality gates, and estimated work for management.
• Ran product demos for the client and interviewed engineering candidates.

Delivered to the client with 91 API endpoints, ~680 tests and a retrieval eval harness with a CI recall gate. A 1,300+ page manual was ingested into 3,800+ searchable chunks.