Search

Every file processed by Roset produces a searchable-index variant (plain text) and chunked embeddings (vector). These power three search modes:

Quick Start#

Search Modes#

Text Search#

Uses Postgres tsvector full-text search on the searchable-index variant. Returns results ranked by ts_rank with highlighted snippets.

• Works immediately after file processing completes
• Best for exact keyword and phrase matching
• Supports pagination via offset

Vector Search#

Embeds your query using OpenAI text-embedding-3-small, then queries Cloudflare Vectorize for the closest chunk vectors.

• Requires an OpenAI provider key configured at /v1/org/provider-keys
• Best for semantic queries where exact words may not appear in the document
• Results include the matched chunk text and chunk index

Hybrid Search#

Runs text and vector search in parallel, then merges results using Reciprocal Rank Fusion (RRF, k=60). This consistently outperforms either mode alone.

• Falls back to text-only if no OpenAI key is configured
• Default mode when mode is omitted

Parameters#

Response#

{
  "results": [
    {
      "fileId": "abc-123",
      "score": 0.034,
      "snippet": "The <b>payment</b> <b>terms</b> are net 30 days..."
    }
  ],
  "total": 42,
  "query": "payment terms",
  "mode": "hybrid"
}

Next Steps#

• Q&A — Ask questions about your files and get answers with citations
• Webhooks — Get notified when new variants are ready for search
• TypeScript SDK — Full SDK reference