Single search endpoint with a mode discriminator.
List of chat messages
1List of repositories to query. Can be strings (slug, display name) or dicts with a 'repository' field.
List of data sources to query. Can be strings (display_name, URL, or source_id) or dicts with 'source_id' or 'identifier' fields
List of local folders to query. Can be strings (display_name or local_folder_id) or dicts with 'local_folder_id' or 'identifier' fields. Local folders are private and user-scoped.
Filter local folder results by classification category (e.g., 'Work', 'Personal')
Filters for local/personal sources (messages, contacts, etc.)
Search mode: 'repositories', 'sources', or 'unified'
Whether to stream the response
Whether to include source texts in the response
Skip LLM processing for faster results (100-500ms vs 2-8s). Set to false for deeper analysis.
Return raw search results without any LLM processing. Returns only sources with scores.
Retrieval strategy: 'vector' (default similarity search), 'tree' (LLM-guided tree navigation for PDFs), or 'hybrid' (both combined)
Maximum tokens in response. Results truncated when budget reached.
100 <= x <= 100000Synthesis model override. Allowed: claude-sonnet-4-5, gpt-5.2-2025-12-11, claude-haiku-4-5-20251001
Minimum similarity for semantic cache hit (non-streaming only)
0.8 <= x <= 1Skip semantic cache (L2) lookup
Generate follow-up questions (adds ~500-1000ms LLM latency). Disabled by default for speed.
Search mode discriminator
"query"Successful Response