Chat
Grounded chat with citations and visible runtime
This route uses the live EigenVertex API. It binds a question to one workspace, runs the compatible answering path for that mode, and exposes the grounded answer with evidence and timing.
Conversation setup
Choose the workspace and active thread
The real value of EigenVertex chat comes from binding the answering flow to a workspace boundary, conversation history and saved messages.
Send a grounded question
Fast path with visible progress
The chat surface now shows a live status and elapsed time while the backend validates the question, searches the workspace and composes the grounded answer.
Fastest path for API-like usage: dense-first retrieval, minimal planning, quick synthesis.
- Dense-first retrieval path optimized for low latency and API-style usage.
- No router, no fusion and no extra sparse exact-match pass.
- Best when you want a quick grounded answer with predictable cost.
Real RAG mode: inject corpus material into the prompt, then let the model synthesize and enrich the answer while keeping source-backed findings explicit.
Direct retrieval + synthesis path, still grounded and inspectable but without the full agent loop.
Qdrant vector retrieval. Effective strategy label: vector.
Lexical exact-term retrieval is part of the retrieval engine today. A dedicated Lexical toggle will be exposed separately.
- Lexical: Postgres full-text recall for exact terms, names, versions and domain-specific vocabulary.
- Vector: indexed chunks in Qdrant, used to collect citations and openable document links.
- Graph: GraphKnowledge relations, useful for connected concepts, methods and related documents.
Moderation is now off by default for speed. Turn it on only when you explicitly want an extra safety review pass around the chat turn.