Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.kb2b.app/llms.txt

Use this file to discover all available pages before exploring further.

Chat is the main way to consume what lives in the POT. You ask in natural language; kb2b pulls the relevant facts from the POT, synthesizes them with Claude and returns a reply with verifiable citations and visible POT Scores.

​
How to start a conversation

  1. Open Chat in the sidebar (/dashboard/chat).
  2. Type your question in the text box. Press Enter to send.
  3. The reply starts appearing in streaming. When it finishes, citations appear below the message.
Conversations are auto-persisted. When you come back to chat the next day, you find the thread where you left it, with every previous message and its citations intact.

​
Anatomy of a reply

Each reply in chat has three visible components: The synthesized text. Claude writes it from the facts SciPot handed over. If you see a number in parentheses in the text (e.g. (2)), that’s a reference to a citation. The citations below the message. They appear as a row of numbered badges (1, 2, 3, …). Each badge shows:
  • The citation index (1, 2, 3, …)
  • The POT Score with two decimals (e.g. 0.92)
  • A level icon classifying the fact (Constitution, Verified, Supported, Inferred, Hypothesis, Speculation)
  • A network icon if the fact has relationships (edges) with other POT facts
The border color of each badge gives you a one-glance read on confidence:
Border colorPOT ScoreQuick read
Green≥ 0.80Verified or Constitution — you can act on this
Yellow0.60 – 0.79Supported — corroborated, reasonably solid
Orange0.40 – 0.59Inferred — the system deduced it, validate before using
Red< 0.40Hypothesis or Speculation — treat as a lead, not as truth
Clicking any badge opens a popover with the fact’s full content, its score, its tags, its edges (relations with other facts) and the retrieval strategy that brought it in (CAG, RAG, GraphRAG). Inside the popover, “Inspect in graph” opens a deeper modal where you can navigate the fact’s connections and understand why it landed in the reply.

​
The streaming UX

While kb2b prepares and synthesizes the reply, you see two indicators:
  • Thinking (three pulsing dots with “Thinking…”) — the system is classifying the question and deciding the retrieval strategy.
  • Reading X (three dots + document or fact name) — the system is reading extra context to enrich the reply.
When the text starts arriving, the indicators disappear and the content renders as streaming markdown. There’s no “blinking cursor” — words simply appear.

​
Which questions work best

Specific questions with a temporal or source anchor. “When does Acme’s contract expire?”, “What did Laura say in the last meeting about pricing?”, “What internal policies do we have about volume discounts?”. These are the ones kb2b answers best because it can anchor the reply to concrete facts. Comparative questions inside the POT’s scope. “How does this account differ from the other three most strategic ones?” works if the POT has material on all four; if not, kb2b tells you data is missing. Questions about what the POT does NOT know. “What key information are we missing to close this account?” — kb2b can list the gaps. Useful before a review meeting.

​
Which questions work poorly

  • Speculation or creativity. “Draft a follow-up email” is not what chat is for. Use a general-purpose LLM for that. kb2b leans toward “what do you know and where from”, not “what can you come up with”.
  • Questions outside the POT’s domain. If the POT is about the Acme account and you ask about your monthly cash flow, it’ll tell you it’s out of scope (assuming scope_exclusion is well-defined in the POT’s Constitution).
  • General-world factual questions. “What’s the capital of Chile?” — kb2b doesn’t answer because no fact in the POT says “Santiago”. For discovery with web context, use Scout.

​
Replies stick around

Every conversation is saved to your workspace’s database. That lets you:
  • Re-open a thread from weeks ago and continue.
  • Share the URL of a specific reply with someone on the team.
  • Audit what was said, with its citations and POT Scores as they were at the time (even if the facts have been updated since).

​
When a reply is important

If a synthesized reply feels so good you want to promote it to a POT fact (with its own POT Score, its relations, its provenance), promotion happens from the Explore and discuss facts screen, not from chat. Chat is the consumption surface; curation is a separate workflow.

​
When a reply has contradictions

Sometimes you’ll see kb2b say something like “the contract expires on 2026-06-30 according to document X, but the last meeting records a negotiation moving the date to 2026-12-31”. That’s a detected contradiction and appears flagged in the Knowledge top banner. Resolve it from Contradictions and resolution — the POT’s average POT Score rises as you lower the pending-contradictions count.
Your plan includes a monthly token quota. Chat consumes tokens based on POT size, question complexity and reply length. If you approach the quota we warn you at 95% and hard-block at 110% to avoid surprises. See Token limits for details.