> ## Documentation Index
> Fetch the complete documentation index at: https://docs.seynlabs.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Chat

> Conversational access to extracted knowledge: streaming answers with citations that resolve all the way down to source records.

Chat is the surface most users live in: ask a question in plain language, get a streamed answer grounded in your organisation's extracted knowledge, with citations on every claim. It's the difference between *"search returned 12 rules"* and *"here's how deal approval works here, and here's the evidence."*

Chat is not a vector-search wrapper around a language model. Answers are assembled from the full [hybrid query pipeline](/platform/query) over knowledge that was extracted, structured, versioned, and human-reviewed before the conversation started. The model writes the prose; the knowledge layer supplies the facts.

## How an answer is produced

```mermaid theme={null}
flowchart LR
    U["User message"] --> R["Hybrid query<br/>(structured + full-text + semantic)"]
    R --> G["Grounded generation"]
    G --> S["Streamed response<br/>+ citations"]
    S --> T["Thread persisted"]
```

1. The message runs through the full query pipeline: the same signals, fusion, and reranking as the API.
2. The model generates an answer **bounded by the retrieved knowledge**, not by its general training. Seyn's value is your organisation's knowledge, not the model's opinions.
3. The response streams token by token, with **citations** linking each claim to the rules it came from. From a citation, the [provenance chain](/platform/provenance) continues down to raw source records.
4. The conversation persists as a **thread**; history is browsable and resumable.

## Grounding and honesty

* **Citations are not decoration.** Every substantive claim carries them, and a citation resolves to a real rule, which resolves to real evidence. An answer you can't drill into is an answer you shouldn't trust.
* **The query bounds the generation.** If the knowledge library doesn't contain the answer, the assistant says so rather than improvising. The failure mode is "I don't have that," not confident fiction.
* **Review status flows through.** Answers built on unreviewed (`inferred`) rules are still answers, but the citations show their status, so a careful reader can weigh them.

## Operational details

| Aspect      | Behaviour                                                           |
| ----------- | ------------------------------------------------------------------- |
| Rate limit  | 50 messages/hour per user                                           |
| Persistence | Threads and messages stored per organisation, with citation records |
| Access      | All roles; chat is the default member surface                       |

## Common mistakes

| Symptom                                                                            | Cause                                                                                      | Fix                                                                                                   |
| ---------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------- |
| "I don't have information about that" for something you know is in the source data | Data was ingested but no extraction run has put it into the active library                 | Walk the [debugging order](/platform/architecture#when-something-looks-wrong-walk-the-chain-in-order) |
| Answers cite rules a human already rejected                                        | The active library version predates the review                                             | Re-run extraction; check citation review statuses                                                     |
| Re-asking the same question gives differently-worded answers                       | Generation is non-deterministic; the *citations* should stay stable even when prose varies | Judge consistency by citations, not phrasing                                                          |
| Treating chat as a calculator or live-data query tool                              | Chat answers from extracted knowledge, not by querying source systems in real time         | Use [pattern metrics](/core-concepts#pattern-metrics) or your BI stack for live aggregates            |

## Related

<CardGroup cols={2}>
  <Card title="Query" icon="magnifying-glass" href="/platform/query">
    The pipeline that decides what chat gets to see.
  </Card>

  <Card title="MCP Server" icon="robot" href="/platform/mcp">
    The same grounded knowledge, for your own agents.
  </Card>
</CardGroup>