RAG Document QA
Turn policies, runbooks, and handbooks into a scoped Q&A pilot: answers are grounded in your agreed documents, citations stay visible, and the FastAPI contract can plug into support or ops workflows.
Commercial pilot shape: agree the corpus, define access and hosting, validate retrieval quality, then hand over a documented API surface — not a generic chat SaaS.
What this solves
Support, ops, and compliance teams store answers in PDFs, wikis, and tickets — but keyword search fails on paraphrases and generic chat tools hallucinate without citations. A VahdetLabs pilot turns your agreed documents into a retrieval corpus, returns answers tied to named passages, and exposes the same behaviour over HTTP for internal tools.
What a pilot produces
From an agreed document set (policies, runbooks, handbooks), you get a repeatable Q&A surface: chunked index, retrieval configuration, cited responses, and integration notes for your network boundary. Outputs suit internal support workflows — not a public open-web assistant. The public reference deployment uses a fixed eight-document sample corpus so stakeholders can verify behaviour before scoping yours.
When this is useful
Repeat policy questions
Teams answer the same SLA, billing, HR, or compliance questions and need faster lookup with traceable evidence.
Cited answers first
You want grounded responses with named passages before rolling LLM access into a wider internal workflow.
Small HTTP contract
Support or ops tools need POST /ask — not a black-box chat widget or open-ended SaaS assistant.
Scoped handover
You prefer a pilot with documented corpus, auth, hosting, and acceptance checks over a broad “AI search” purchase.
How the system works
End-to-end path from agreed documents to cited answers over HTTP:
POST /ask
Indexing runs offline; the API loads FAISS + metadata at startup and keeps the online contract stable while you refresh documents on a schedule.
-
1
Scope the corpus
Agree policies, runbooks, or handbooks — the retrieval boundary, not the open web.
-
2
Chunk, embed, index
Chunk with overlap, embed with a fixed model, build FAISS — reproducible offline pipeline.
-
3
Retrieve + generate with citations
Top-k passages ground the LLM; every answer lists document, page, and chunk id.
-
4
Integrate via FastAPI
POST /askfor support tools, scripts, or internal widgets — same contract as the reference deployment.
Acceptance checks focus on retrieval hit rate and source alignment, not vanity chat UX.
Where to try it
This overview lives on the VahdetLabs project page (rag.vahdetlabs.com). The reference API host and interactive demo use a fixed sample corpus for evaluation — an engineering reference, not a consumer landing experience.
Scope and limits
- Not a hosted multi-tenant chat product or open-web research agent.
- Answer quality depends on corpus quality, chunking, and retrieval — pilots define acceptance checks.
- Production work includes auth, network boundary, and corpus governance beyond this reference UI.