Skip to main content
Tremor’s AI assistant translates natural-language questions into executable SQL. The service runs on Anthropic Claude with Tremor-specific tooling for schema discovery and validation.

​
Choose the Right Mode

Query Writer

Default mode focused on generating optimized SQL. Ideal for analysts who want a ready-to-run statement.

Analyst

Adds written commentary and requires an Exa API key for external research augmentation.
Select the mode by setting the mode field in the request body (query_writer or analyst).

​
Streaming Integration

POST /api/assistant/generate/stream emits Server-Sent Events (SSE) so you can render incremental updates: 
curl -N -X POST https://tremor.sh/api/assistant/generate/stream \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "message": "Show the highest-volume AI regulation markets this week",
    "mode": "query_writer"
  }'
Each event follows the format: 
data: {"type":"plan","payload":{"step":"Reviewing schema"}}

data: {"type":"sql","payload":{"query":"SELECT ..."}}

data: {"type":"final","payload":{"query_id":"...","sql":"SELECT ..."}}
Listen for type=final to retrieve the finished SQL. Handle type=error as a terminal state.

​
Guardrails & Validation

Behind the scenes the assistant calls:
  • GET /api/tables and /api/table/{namespace}/{table} for schema metadata
  • POST /api/query/validate to confirm syntax
  • Optional semantic search utilities when your prompt references similar markets
This orchestration ensures generated statements are executable and scoped to Tremor’s datasets.

​
Rate Limits & Cost Awareness

  • 20 assistant requests per minute per API key
  • Streaming responses return within ~3–6 seconds for typical prompts
  • Claude usage is metered on Tremor’s side, so high-frequency automation should cache outputs when possible

​
Productive Workflows

  • Explain existing SQL: add conversation history containing the original query so the assistant can annotate or optimize it.
  • Iterate on prompts: supply previous turns via conversation_history to refine results without losing context.
  • Auto-execution: chain the assistant with /api/query once a final event arrives to deliver charts or dashboards on the fly.
See the API reference entry for exact response fields, or pair the assistant with query recipes when you need deterministic baselines.