Built-in Features & Database
The engine provides built-in tools and call management capabilities out of the box. Enable or disable them via [tools] builtin in chakra.toml.
1. Built-in Tools
| Built-in tool | Parameters | What it does |
|---|---|---|
save_caller_memory |
summary, language_preference |
Stores a short summary of the call against the caller's phone number |
capture_feedback |
score (1 to 5) |
Records caller satisfaction rating |
query_knowledge_base |
search_query |
Searches data/ documents and returns top grounded passages |
end_call |
None | Says goodbye and hangs up. Requires tools.end_call.description in instructions.json |
2. Caller Memory & Greetings
When [features] database = true is set and a PostgreSQL database is connected, the engine automatically tracks callers:
- Returning Callers: A caller is identified by phone number. If they have called before, the agent greets them using greeting.returning_caller_template, which can reference {last_date}, {base_greeting}, or any stored memory fields.
- Memory Fallback: If a call ends abruptly before the model calls save_caller_memory, the engine automatically saves an extractive summary from the caller's spoken turns ([features] memory_fallback = true).
3. Silence Handling
When the caller goes quiet, the agent checks in using prompts from [silence] prompts in order. If the caller still does not respond, it speaks [silence] goodbye and ends the call.
- monitor mode: Used on the turn-based chakra pipeline.
- away mode: Used on Gemini Live (user_away_timeout_seconds, max_check_ins).
4. Maximum Call Length
Set [call] max_duration_seconds to cap long calls (e.g., 900 for 15 minutes). The agent speaks the line at max_duration_goodbye and closes the session. Set to 0 for unlimited duration.
5. Daily Talk Time Per Caller
Daily talk time limits are configured centrally in Chakra Console and enforced per phone number:
1. When a call connects, the engine sums up the caller's total talk time since midnight in the configured timezone (Asia/Colombo).
2. If less than min_usable_seconds remains, the call is rejected before the greeting with call_limit.limit_reached.
3. If the quota approaches during a call, the caller is warned (call_limit.approaching_limit) and disconnected when time runs out (call_limit.time_up).
4. Daily talk limits reset automatically at midnight.
6. Call Records & PostgreSQL Schema
When DATABASE_URL is configured, the engine maintains the following tables:
| Table | One row per | Main columns |
|---|---|---|
call_sessions |
Call session | session_id, phone_number, language, status, duration_seconds, start_time, end_time, handoff_requested, location fields, session_type |
caller_memory |
Caller profile | phone_number, summary, language_preference, last_interaction_date |
session_feedback |
User rating | session_id, rating, comment, timestamp |
conversation_turns |
Spoken turn | Transcripts of each turn with speaker role and timestamp ([features] transcripts = true) |
knowledge_gaps |
Unanswered query | Questions the agent could not answer, frequency count, and timestamp |
system_errors |
Error | Unhandled exceptions and system warnings for diagnostic auditing |
Data Privacy: This data stays entirely in your own PostgreSQL database. Chakra Labs receives only aggregate call duration, tokens, and billing usage metrics.
7. Distributed Tracing with Langfuse
Set the standard Langfuse environment variables in .env to record end-to-end tracing:
- LANGFUSE_PUBLIC_KEY
- LANGFUSE_SECRET_KEY
- LANGFUSE_HOST
Every call generates a trace containing each turn, token usage, latency breakdowns, and tool invocations.