Early access — in design with pilot teams

Your vector database executes queries brilliantly. Someone still has to construct the right ones.

If you run search inside an AI-native SaaS product, you are probably hand-maintaining the layer around TurboPuffer or Pinecone: rewriting user questions into multiple retrieval queries, wiring structured and permission filters, fusing and reranking results, and guessing whether any of it actually worked.

The LemonCrow Search Compiler turns a natural-language request into a typed, validated, permission-constrained search plan — then executes it, ranks the results, and tells you whether the search worked.

You send

{
  "query": "Candidates who passed interviews but never received an offer",
  "principal": { "user_id": "recruiter_123", "tenant_id": "acme" },
  "schema": "recruiting_v1"
}

LemonCrow compiles

{
  "entity": "candidate",
  "conditions": [
    { "relationship": "interviews", "where": { "result": ["pass", "strong_pass"] } },
    { "relationship": "offers", "exists": false }
  ],
  "permission_filter": { "accessible_job_ids": ["job_1", "job_2"] },
  "retrieval": [
    { "type": "structured" },
    { "type": "bm25" },
    { "type": "vector" }
  ],
  "ranking_profile": "candidate_search_v3"
}

Not an agent, not a RAG platform. Its job ends after returning correctly ranked records.
Your authorization stays yours. It consumes permission filters from your existing system (or Cerbos / Oso) — it never becomes the policy engine.
Every plan is testable. Real questions, expected records, generated plans, and failures are stored and replayable, so a schema or prompt change can't silently degrade search quality.

We already built this once — for code

LemonCrow's shipped product does exactly this for codebases. Before: agents grep with hand-tuned regexes, re-read files, and loop until something matches. After: one code_search call takes plain keywords or a question and returns the right symbols, ranked source, and their connections — query construction, filtering, and ranking compiled behind one call.

That layer is benchmarked continuously (MRR-tracked retrieval, matched SWE-bench runs). The Search Compiler applies the same discipline — typed plans, replayable evals — to your product's data instead of code.

We're picking a handful of pilot teams

If your team maintains hand-written query construction around a search database, we want 30 minutes to hear exactly where it breaks. No pitch — the answers shape what gets built.

Book a 30-minute call