Search notes
Performs hybrid vector + full-text search and returns results grouped by
source note with matching text snippets. search_mode=comprehensive
additionally reranks results with a cross-encoder against person/company
context (not just raw note text) for broad, ambiguous asks — e.g. “give me
everyone I’ve researched who fits this profile” — at the cost of extra
latency (~4-5s) and a small per-request cost. Every other mode stays on
the fast, cheap path for direct/precise lookups.
Scope to a specific entity/opportunity with entity_uuid+entity_type, or
across many opportunities at once with opportunity_uuids/
active_opportunities_only — e.g. “data-room concerns across all active
opportunities this quarter” resolves as one call scoped to open
opportunities, no per-note follow-up needed. Each result’s associations
field carries its linked entities/opportunities inline for client-side
grouping.
Authorizations
Body
Search query text.
Restrict to notes associated with an open/active-stage opportunity. Combines with opportunity_uuids (intersection) when both are given, or applies tenant-wide when omitted.
ISO datetime upper bound on note creation date.
Type of entity_uuid (PERSON/COMPANY/OPPORTUNITY). Must be paired with entity_uuid.
PERSON, COMPANY, OPPORTUNITY Scope results to notes associated with this single entity. Must be paired with entity_type.
Max number of notes to return (1-200).
Scope results to notes associated with any of these opportunities — use for asks spanning multiple opportunities at once, e.g. 'data-room concerns across all active opportunities this quarter'. At most 200 UUIDs.
Controls how results are ranked. Defaults to 'recency' when omitted. Choose based on the user's intent:
- 'recency': favours recent notes with a time-decay score boost. Use for: 'latest notes', 'recent conversations', 'what happened last week'.
- 'historical': favours older notes. Use for: 'first meeting', 'earliest notes', 'how did we start with X'.
- 'comprehensive': widens the candidate pool AND reranks results with a cross-encoder scored against each note's associated person/company context (job title, company description, mentions), not just raw note text — use for broad/ambiguous asks like 'give me everyone I've researched who fits this profile' or 'all notes mentioning Y'. Slower (~4-5s) and has a small per-request cost; every other mode stays on the fast, cheap path, so prefer this only for genuinely broad asks, not direct/precise lookups.
- 'balanced': distributes results evenly across time. Use when a representative sample across full history is needed.
- 'relevance': pure semantic match, no time factor. Use when recency is explicitly irrelevant.
recency, historical, relevance, comprehensive, balanced ISO datetime lower bound on note creation date.