Skip to content

07-chat-threads.md

Latest commit

 

History

History
182 lines (124 loc) · 7.35 KB

07-chat-threads.md

File metadata and controls

182 lines (124 loc) · 7.35 KB
Raw

07 Chat Threads

What chats are in this system

Chats are persistent conversations against the wiki. Each chat is a markdown file in chats/. Folders are real directories. A chat can be opened, continued, archived, deleted, or promoted into a permanent wiki page.

The distinction from one-off queries: chats persist across sessions and accumulate context across many turns. The query mode is for "ask once and move on."

File format

One file per chat. Filename: YYYY-MM-DD-HHMM-{slug}.md.

---
id: 0193abc12345
title: Comparing error correction codes
folder: deep-dives
created: 2026-05-23T14:30:00Z
updated: 2026-05-23T15:12:00Z
model: anthropic/claude-3-5-sonnet
pinned: false
message_count: 6
tags: [quantum, error-correction]
---

## user [14:30:00]
What error correction schemes are mentioned in my sources?

## assistant [14:30:15]
Three are referenced in the wiki: [[surface-codes]] (Shor 1995), [[stabilizer-codes]] (Gottesman thesis), and [[topological-codes]] (Kitaev 2003).

Each has tradeoffs. [[surface-codes]] are the most studied and have the most experimental progress...

## user [14:35:00]
Can you compare them in a table?

## assistant [14:35:45]
| Code type | Threshold | Overhead | Experimental progress |
|-----------|-----------|----------|----------------------|
| [[surface-codes]] | ~1% | High | Most |
| [[stabilizer-codes]] | varies | Medium | Moderate |
| [[topological-codes]] | ~1% (theoretical) | Lower | Early |

The thresholds and overheads are approximate; see [[error-correction-thresholds]] for the latest measurements.

## user [14:40:00]
Promote the comparison to a wiki page.

## assistant [14:40:30]
Done. Created [[error-correction-codes-comparison]] with the table and context. Linked from [[error-correction-thresholds]] and from each individual code page.

Folder organization

Folders are filesystem directories under chats/. Default folders created on first launch:

chats/
├── inbox/       # default for new chats
├── pinned/      # convention for important chats
└── archive/     # convention for older chats

The user can create any folder by typing a folder name in the move dialog. Nested folders are allowed but discouraged in V1 (one level deep only).

A chat is in whatever folder it physically lives in. Moving = fs.rename. No abstraction.

Operations

Create a chat

User clicks "New chat" or starts typing in the chat composer. We:

  1. Generate a UUID and timestamp-based filename
  2. Create file in chats/inbox/ with empty messages array
  3. Insert row in chats SQLite table
  4. Navigate to the chat view

Send a message

  1. Append ## user [HH:MM:SS] block and message text to the file
  2. Build context for the LLM:
    • System prompt: "You are answering questions in a persistent chat thread against an LLM wiki. The user has access to the full wiki content. Cite pages with [[slug]]. Reference earlier messages when relevant."
    • Wiki index
    • Relevant pages (top K based on the latest user message)
    • Full chat history (truncate older messages if context limit approaches)
  3. Stream the response, writing to a buffer
  4. Append ## assistant [HH:MM:SS] block with the response to the file
  5. Update updated timestamp and message_count in SQLite

Rename a chat

  1. Update title in frontmatter (file content)
  2. Update title in SQLite
  3. Filename stays the same (we keep timestamp-based filenames for sort stability)

Move a chat to another folder

  1. fs.rename the file from chats/{old}/file.md to chats/{new}/file.md
  2. Update folder in frontmatter
  3. Update folder in SQLite

Pin a chat

Set pinned: true in frontmatter. Update SQLite. Pinned chats sort to the top of their folder.

Delete a chat

Move file to .llm-wiki/trash/chats/. Delete row from SQLite. Trash is purged after 30 days (background task on startup).

Promote an assistant message to a wiki page

This is the feature that makes chats valuable for knowledge accumulation. User clicks "Save as wiki page" on any assistant message. We:

  1. Show a modal: "Create a wiki page from this answer?"
  2. Pre-fill: title, slug, type (let user pick), tags
  3. On confirm: extract the message text, write to wiki/{slug}.md, update index, update SQLite
  4. Add a small "→ saved to [[slug]]" annotation in the chat file after the promoted message

Ingest the whole chat as a source

For a chat that contains a lot of useful synthesis, the user can ingest the entire chat as a source. This runs the same ingest pipeline as for a PDF or article:

  • The chat file is treated as the source
  • The LLM extracts entities, concepts, decisions discussed
  • Creates and updates wiki pages
  • The chat stays where it was; ingestion is non-destructive

UI views

Chat list (sidebar or dedicated view)

[+ New chat]                                    [search]

PINNED
  ★ Architecture decisions                      2 days ago
  ★ Open questions                              5 days ago

INBOX                                           [collapse ▼]
  ◯ Error correction comparison                 1h ago
  ◯ Shor vs Grover                              3h ago

DEEP DIVES                                      [collapse ▼]
  ◯ Quantum supremacy claim review              yesterday
  ...

ARCHIVE                                         [collapse ▶]

Each entry shows: pinned star, title, last-updated relative time. Right-click or hover for menu (rename, move, archive, delete, pin).

Chat view

Asymmetric Claude-style layout (shipped 2026-05-24, commit 087a51a). User messages render as right-aligned bubbles (max-w-[80%], filled bg-primary/10, rounded with a squared top-right corner). Assistant messages flow left-aligned at max-w-[92%] with no bubble — long cited answers, tables, and code blocks need width to breathe, so we don't constrain them. Position alone tells you who said what; role badges dropped.

Messages render as markdown with [[wikilinks]] clickable. Composer at the bottom. Header shows the chat title (editable inline), folder dropdown for moving, Pin / Ingest → wiki / Delete actions, and the model the chat uses (set on creation, stored in frontmatter).

Each assistant message has a small action row beneath: Save as wiki page (opens the promote dialog) plus the timestamp. Whole-chat Ingest → wiki lives in the header so a useful thread can be filed back into the permanent wiki layer with one click — per Karpathy's "good answers can be filed back into the wiki as new pages".

Context window management

Chats can grow long. Strategy:

  • Always include: system prompt, wiki index, top-K relevant pages
  • Include all messages until we hit 70% of model context
  • If we go over: keep the most recent 10 turns verbatim, summarize earlier messages into a compressed system note
  • Show the user a banner when summarization kicks in

Search across chats

Use SQLite FTS5 on chat content. Index every message body. Allow filters: by folder, by date range, by model used. Click a result to open the chat at that message.

What chats are NOT in V1

  • No multi-user chats (single user only)
  • No real-time collaboration
  • No voice or audio messages
  • No chat templates
  • No automatic chat naming (user names their own; we generate a default from the first user message)
  • No embeddings-based "related chats" feature
  • No chat folder sharing or export (the files are markdown, user can copy them manually)