# Search customer feedback > Find related comments through semantic REST search and understand keyword fallback and MCP search. Canonical URL: https://pollenate.dev/docs/ai/semantic-search/ ## Search comments by meaning The dashboard search and `POST /organizations/ORG_ID/search` can find semantically related inbox comments. For example, “trouble inviting a teammate” may find comments that describe a confusing member-invite flow without using those exact words. ```bash curl https://api.pollenate.dev/organizations/ORG_ID/search \ -H "X-Pollenate-Key: $POLLENATE_READ_KEY" \ -H "Content-Type: application/json" \ --data '{ "query": "trouble inviting a teammate", "inboxId": "INBOX_ID", "limit": 20 }' ``` Use an organization-wide organization key with `read`. The body requires a query of 1–500 characters. Optional `brandId` and `inboxId` restrict the search; `limit` defaults to 20 and accepts 1–100. ```json { "query": "trouble inviting a teammate", "mode": "semantic", "results": [], "total": 0 } ``` This illustrates the envelope; actual mode and rows depend on the available index. Semantic rows include a `relevance` value alongside the event ID, comment, score, type, context and inbox/brand names. Relevance is a similarity score, not a probability that an answer is correct. ## Know what is indexed The embedding process indexes nonempty **inbox feedback comments**. It skips events without comments. Arbitrary context and all survey answers are not automatically embedded as one searchable corpus. When semantic results are unavailable or empty, the REST route falls back to substring matching across comment and serialized context, returning `mode: "text"`. Text fallback sorts newest first and does not add semantic relevance scores. Newly collected comments may take time to appear in the semantic index because indexing runs after ingestion. Search for an exact phrase to check the fallback, or read the [feedback event](/docs/api/feedback/) directly to verify persistence. ## MCP search is a different contract The current `search_feedback` MCP tool searches keywords in comments and context. It is not the REST semantic-search implementation. The tool accepts `query`, optional `brandId`/`inboxId`, and `limit` (default 20, maximum 100). A person's results remain constrained by their current role and brand access. Searches through REST count against the organization's daily search allowance. User-backed MCP searches also count against that allowance; the existing organization-key MCP path has different accounting. A daily-limit response includes reset timing. Changing the query repeatedly will not bypass the limit. ## Search docs instead of private feedback If you are looking for an endpoint, installation step or permission, use [documentation search](/docs/api/documentation-search/). That endpoint searches public documentation and never queries your customers' feedback.