Knowledge bases — the RAG Studio

A base = an isolated workspace of your documents, queryable with exact citations. The studio (Knowledge bases menu) offers: Ask · Documents · Chunks · Graph · Settings · Cost simulator.

Ingestion

PDF, Word, PowerPoint, Excel, images, Markdown/TXT/HTML/CSV — including illustrated PDFs (figures and pages preserved: vision mode). Re-uploading a modified file cleanly replaces it.

Asking

• Modes: ✦ Auto (planner decides) · Quick · Full · 🎯 Deep (reranking) · ◈ Graph (multi-hop) · ❖ Synthesis (whole corpus).
• ⚡ Agent: decomposes multi-part questions.
• Scope: whole corpus or a single document (selector above the chat).
• Cited [N] answers, clickable → exact passage, PDF page, extracted figures.
• Conversations: every exchange is stored; resume any of them (🕘 History).
• Transparent cost: latency, chunks, tokens and $ per query; simulator before onboarding a corpus.

Per-base settings

Retrieval profile (generic · legal · enterprise · timeline · catalogue), analysis language (FR/AR/EN…), domain entities, multimodal (image extraction: off / pages / per figure), dedicated LLM & OCR models.

Sharing & teams

Public (link) or email sharing; project assignment; enterprise team grants. Bases shared with you appear 🌐 in every selector across the platform (chat, applications).

Via API

POST /api/v1/kb                       # create (profile, language, entities…)
POST /api/v1/kb/{id}/documents        # ingest (base64) — poll status
POST /api/v1/kb/query/stream          # SSE: meta · source(+images) · token · cost · done
GET  /api/v1/kb/{id}/graph|chunks|entities|config

And in chat: pass kb_ids so ANY agent grounds itself on your bases (API).