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

# Ask

> Chat with a single meeting or across your whole history — local retrieval, your own model, streamed answers.

**Ask** turns your transcripts into a conversation. It runs locally: retrieval and embeddings are on-device, and only your own model call (the Anthropic or OpenAI key you set for notes) leaves your machine. Answers stream in token-by-token.

## Two ways to ask

<CardGroup cols={2}>
  <Card title="This meeting" icon="message">
    The **Ask** tab on any session answers from that call's transcript. Good for "how did I do?", "what did I commit to?", or any detail from the conversation.
  </Card>

  <Card title="All meetings" icon="layer-group">
    **Ask across all meetings** (from the dashboard header) searches your whole history and cites the sessions it drew from. Click a source to open it.
  </Card>
</CardGroup>

## A thinking partner, not a search box

Ask reads the transcript and talks it through with you — sentiment, what landed, what you may have missed — and answers direct factual questions from what was actually said. It won't invent names, numbers, or dates that aren't in the transcript; if something isn't there, it says so.

Follow-ups build on earlier turns. Long chats keep their context: older turns are folded into a running summary (never dropped) so the thread stays coherent without exceeding the model's context window.

## Saved conversations

* **Per meeting:** each session keeps a single Ask chat, restored when you reopen it.
* **Across all meetings:** keep as many chats as you want — **New chat**, switch, **rename**, or **delete** from the sidebar. Each is auto-titled from your first question and saved locally.

Answers stream as they generate, and the input stays pinned to the bottom like a normal chat. If you navigate away mid-answer, the turn isn't lost — generation finishes in the background and the full answer is saved.

## How retrieval works (across all meetings)

<Steps>
  <Step title="Index (on-device)">
    Every session's transcript is chunked and embedded locally with **`Xenova/all-MiniLM-L6-v2`**. New sessions are indexed automatically; older ones are backfilled the first time you open Ask.
  </Step>

  <Step title="Match">
    Your question is embedded and ranked against the chunks by cosine similarity.
  </Step>

  <Step title="Answer">
    The top matches (within a token budget) plus recent conversation go to your model, which answers with inline citations like `[1]`.
  </Step>
</Steps>

The per-meeting Ask tab skips retrieval — a single call's transcript usually fits, so it's fed directly for the most accurate read.

## Privacy

Chunks, embeddings, and saved chats live in **SQLite** on your machine. There is no Raven server and no cross-account sync. Deleting a session also removes its transcript chunks and its Ask chat.