Retrieval-augmented generation over your internal data: hybrid search, permission-aware retrieval, citation-backed answers, and evaluation you can measure.
The demo answered three questions beautifully
Naive RAG works impressively on a curated sample and degrades badly on a real corpus. Pure vector search misses exact matches like part numbers, error codes, and contract clauses, because semantic similarity is the wrong tool for identifiers. Chunking splits tables and procedures mid-thought so retrieved context is incoherent. Permissions get ignored, so the assistant cheerfully surfaces the salary review to whoever asks. And with no evaluation set, nobody can tell whether last week's prompt change helped or quietly broke retrieval.
What we build
- Ingestion across documents, wikis, tickets, email, and databases with incremental re-indexing
- Hybrid retrieval combining vector similarity with keyword search, so exact identifiers are still findable
- Structure-aware chunking that keeps tables, procedures, and clauses intact
- Permission-aware retrieval enforced at query time against your existing access model
- Citation-backed answers linking to the source passage, so users can verify rather than trust
- An evaluation set and retrieval metrics so every change is measured, plus honest refusal when the corpus lacks the answer
How we work
Collect real questions from the people who will use it, and build the evaluation set from those
Ingest a bounded, high-value corpus first rather than everything at once
Measure retrieval quality before tuning generation, since most failures are retrieval failures
Pilot with a real user group, review the questions it fails, and expand deliberately
Typical stack
Frequently asked questions
They solve different problems. RAG changes what the model knows and is the right answer for factual, changing information: update a document and the next answer reflects it, with a citation. Fine-tuning changes how a model behaves, such as tone, format, or a specialized task. Fine-tuning on facts is expensive, goes stale immediately, and cannot cite sources.
Filtering happens at retrieval time against your existing access model, so a user's query can only ever match documents they are already entitled to read. Permissions are never handled by asking the model to withhold information, because that is not a security boundary.
Three things together. Answers are constrained to retrieved context, every claim carries a citation the user can open, and the system is built to say it does not know when retrieval returns nothing relevant. That last behavior is the one teams skip and the one that determines whether people keep trusting it.
With a labelled evaluation set built from real user questions. We track retrieval metrics such as whether the correct passage appeared in the top results, alongside answer correctness. Without this, tuning is guesswork and regressions go unnoticed for weeks.