Core
AI Knowledge Copilot
A grounded document assistant — retrieval before generation, visible excerpts, confidence signals. Not a generic chatbot.
- Problem
- Internal knowledge is fragmented. Search returns noise; an LLM without retrieval answers fluently and wrong.
- Why it matters
- This is the pattern Ask My Lab must follow. Retrieval, citations and explicit uncertainty before any answer.
Stack: Python · React · RAG
Architecture
File ingest → chunks → retrieval → context narrowing → generation → excerpts, confidence and a “why this answer” panel.
Product modes: Ask, single-document summary, compare, cross-document synthesis, metadata filters, session history.
French-first UI, FR/EN switch.
Data
PDF, DOCX, TXT, MD, CSV. No OCR on scanned PDFs. No XLSX, PPTX, email or connectors yet.
Approach
- Retrieval before generation.
- Source narrowing before the final answer.
- Confidence signals instead of false certainty.
- Workflows (summary, compare, synthesis) rather than one chat box.
Results
Evidence: architecture, workflows and the trust layer in the public repo. No published retrieval benchmark numbers here.
Limitations
No auth, no workspace boundaries, no permissions. No OCR. An internal demonstration product, not a client SaaS.
Learnings
“Grounded” has to show up in the UI: excerpts, confidence, document compare. Ask My Lab will reuse that discipline, with a 30+ question set from V3.