Batch Extraction

Batch Extraction

When you need to extract state for multiple messages at once — either catching up after chatting with auto-extract off, or re-extracting an entire chat with new settings.

Extract Remaining

/bt-extract-remaining

Finds the last message that has extracted state and runs extraction for every unprocessed message after it.

When to Use

  • You’ve been chatting with auto-extract off and want to catch up
  • You started a chat before installing BlazeTracker and want to extract existing messages
  • Extraction was interrupted (browser crash, abort) and you want to resume

How It Works

  1. Scans messages to find the last one with extracted events
  2. If no extraction exists, runs initial extraction on the first message
  3. Runs event extraction for each subsequent unprocessed message in order
  4. Shows progress as each message is processed

What If Early State Is Wrong?

Since each message’s extraction builds on previous state, errors compound. If message 3’s location is wrong, every subsequent message will reference the wrong location.

Fix: Run /bt-extract-remaining, then correct errors at the earliest wrong message using the event editor. Downstream state updates automatically.

Extract All

/bt-extract-all

Clears all existing state and re-extracts the entire chat from scratch.

When to Use

  • You’ve tuned your prompts and want to see the results across the whole chat
  • State has drifted significantly and corrections aren’t practical
  • You want a clean slate after changing track settings

How It Works

  1. Deletes all events, snapshots, and state from the event store
  2. Runs initial extraction on the first message (creates a new snapshot)
  3. Runs event extraction for every subsequent message in order
  4. Shows progress as each message is processed
This operation is slow for long chats. Each message requires a full extraction cycle (5-20+ LLM calls depending on configuration). A 100-message chat could take 10-30 minutes.

Reducing Extraction Time

  • Disable unnecessary modules in Settings > Tracking before running. Each disabled module is one fewer LLM call per message.
  • Enable prefix caching on your backend (see Setup)
  • Use a faster model if accuracy is less critical for the batch run

Aborting

During either batch operation, you can abort by clicking the stop button in SillyTavern.

When aborted:

  • Events from completed messages are kept
  • The current message’s extraction is discarded (all-or-nothing per message)
  • You can resume later with /bt-extract-remaining — it picks up where you left off

Progress Tracking

During batch extraction, the UI shows:

  • Which message is currently being processed
  • Which extraction phase is running (time, location, characters, etc.)
  • The total number of messages remaining