Upload a handbook, wiki, or SOP. New hires ask questions and receive cited answers grounded in the source documents. Shipped and live.
Ask anything about your company docs, policies, and processes.
Rebuilt snapshots of the live product. Each shows a specific stage of the retrieval loop.
In production this is white-labeled to a single company's knowledge base. The public demo offers three indexed corpora so a visitor can test the retrieval loop without uploading anything.
Ask anything about your company docs, policies, and processes.
These are real documents we've pre-loaded for you. Pick one to get started.
GitLab's public company handbook — PTO, benefits, values, spending policies, onboarding
3070 chunksSBA.gov guides on small business taxes — structures, deductions, deadlines, filing requirements
100 chunksSupabase official docs — database, auth, storage, edge functions, migrations
109 chunksOnce a corpus is selected, the chat opens empty with four contextual starter questions drawn from the active knowledge base. One click starts a query.
Ask anything about your company docs, policies, and processes.
The generator prompt constrains the LLM to retrieved chunks and requires inline source labels. Output is a numbered response with Source N tags next to each claim.
Each source card displays the retrieved text, cosine relevance score, filename and section, and a link to the document viewer.
The viewer opens the source document inline with the retrieved chunk highlighted. Surrounding context remains visible for verification.
Once you are certain that you are eligible to work in your requested location country, you will need to provide proof of eligibility by [uploading the proof in the Documents tab of your Workday profile](https://docs.google.com/document/d/1
If you move before establishing eligibility to work, but then cannot establish eligibility to work in the new location within the first six months of residency, it may affect your ability to continue your role with GitLab.
Review the Country hiring guidelines to see if we can support your relocation.
*If you are moving to a different country, please start this process no later than 3 months prior to your ideal relocation date.
Vite + React frontend. FastAPI backend with four routers under /api. ChromaDB persists one collection per topic. OpenAI text-embedding-3-small for vectors, gpt-4o-mini for generation.
Chunking splits on paragraph → sentence → word boundaries with 50-character overlap and header detection. Retrieval returns top-5 chunks filtered at cosine ≥ 0.3. The generator prompt restricts the LLM to retrieved context and requires inline source tags.
Pre-load real corpora. An empty RAG demo doesn't demonstrate retrieval. Three indexed knowledge bases (GitLab, SBA, Supabase) let a visitor test the loop immediately.
Collapse sources by default. The answer is primary; citations sit below, collapsed. Users who trust the answer move on. Users who need verification expand.
Display chunks verbatim. Retrieval sometimes returns fragments with markdown or hashes. The source card shows them as-is so users can diagnose why an answer was produced.