Context Repair & Rolling Summary

active-call provides advanced context management features to handle complex conversational scenarios.

Context Repair

This feature addresses the "fragmentation" issue caused by pauses in speech. For example:

1. User: "My car broke down..."
2. (Pause)
3. Bot: "Where are you?"
4. User: "...at the station."

Without context repair, the bot would likely repeat the question "Where are you?". With context repair, the system detects that the user's second statement is a continuation of the first, removes the bot's interruption, and merges the user's messages. The LLM sees: User: "My car broke down... at the station."

Configuration

Enable this feature by adding "context_repair" to the features list in your playbook llm config.

---
llm:
  features: 
    - "context_repair"
  repair_window_ms: 3000 # Window to detect continuation (ms), default: 3000
---

Rolling Summary

For long conversations, this feature automatically summarizes the history to prevent token overflow while maintaining context.

Configuration

Enable this feature by adding "rolling_summary" to the features list.

---
llm:
  features: 
    - "rolling_summary"
  summary_limit: 20 # Number of messages before triggering summary, default: 20
---