Skip to main content
Semantic search is opt-in. Existing keyword searches and the Activities page keep their current behavior. Use CLI 0.3.0 or newer:
Browser login remains the default. Add --api to use your separate API-key allowance. --pretty changes JSON formatting only.

​
Filters and ranking

Combine --semantic with company, location, industry, headcount, date, hiring, funding and advanced group filters. Add --q when the activity title must also contain specific literal keywords. Use the country value returned by search results (for example US) rather than substituting a display label.
The query is embedded with the same Gemini gemini-embedding-001, CLUSTERING, 768-dimensional configuration used for the published corpus. Raw vectors are not returned. Search retrieves up to 1,000 approximate nearest published embedded signals, then applies structured and keyword filters. Filters do not search outside this bounded pool. Narrow filters may therefore return fewer results even when other matching records exist elsewhere. Signals without embeddings are not included. Items are ordered by cosine distance ascending (equivalently, similarity descending), with descending signal ID as the tie-breaker. similarity ranges from -1 to 1. It is a ranking score, not a probability, confidence threshold or guarantee of business fit. Review the signal evidence and dates before making a recommendation. No results are synthesized. The response retains items and nextCursor, with semantic metadata under search:

​
Limits and pagination

  • Semantic query: 3–1,000 characters.
  • --limit: 1–50, default 50.
  • Only returned records count toward usage. A limit of five can return fewer than five.
  • Continue with --cursor, repeating the same semantic query, filters and limit. Pages are never fetched automatically.
  • Cursors expire after one hour. They bind the query, filters, limit, last rank and first-page publication cutoff.
  • --page, --ids and --count / countOnly=true are not supported in semantic mode. Keyword searches retain these options.
Approximate nearest-neighbor retrieval and corpus corrections can change candidate coverage between calls. The publication cutoff excludes later publications, but this is not a frozen database snapshot or an exhaustive export.

​
API

Browser-login CLI uses /api/cli/v1/signals with its existing CLI session. Do not send browser cookies to API-key endpoints.

​
Billing, privacy and failures

Semantic and keyword modes use the same record allowances: browser-login CLI uses CLI records; API keys use separate API records. One returned signal consumes one record. Empty results, failed requests and idempotent replays consume no data records. There is no separate SeedOps semantic-record surcharge. The semantic query text is sent to Google’s existing Gemini embedding service. Keyword-only searches do not use that service. A bounded in-memory cache may reuse an identical query embedding for one hour; this reuse is best-effort across Worker instances. Authentication tokens, API keys and full company profiles are not sent for query embedding. Missing embedding configuration or provider/database failures return 503 with an actionable error. Semantic mode never silently falls back to keywords. Retry with the same --idempotency-key to avoid charging twice for a completed request. Do not repeatedly retry 429; respect Retry-After. Use seedops usage to check your allowance for free.