Service · RAG
Search across your files that cites a source for every answer, scored against a set of real questions so you know the recall before launch.
EngineerOS scores retrieval on 20 typed queries over a corpus where 14 of 22 documents are deliberate distractors. The first, easier corpus scored 100% for every strategy, which meant it measured nothing.
Proof: EngineerOSEqual-weight hybrid search scored worse than keyword alone until the semantic leg was down-weighted to 0.35. The weights come from the eval, not from a default.
Proof: EngineerOSRegression tests assert the recall floors in CI, so a change to chunking, embeddings or prompts that hurts retrieval cannot ship unnoticed.
Proof: EngineerOSAnswers carry the passage they came from. The research agent's pipeline ends in a report where every claim is cited.
Proof: ArchMind Research AgentBefore launch we write a set of real questions with known answers from your documents, and the system is scored on them. The score is a number in the handoff, and a test keeps it from dropping later.
Usually both, weighted by what the evaluation shows. On EngineerOS keyword search beat an equal-weight hybrid until the semantic side was down-weighted, so the mix is measured per corpus rather than assumed.
In your own database and cloud. The default stack is PostgreSQL with pgvector, which keeps documents, embeddings and permissions in one place you already control.
Two to four weeks for a scoped first version. Projects are priced from the published bands, ₹15,000–35,000 for a single workflow and ₹25,000–60,000+ for a full product, and the exact price is fixed in writing after scoping.
Two or three sentences about the problem is enough. I reply within one working day, and the first 15-minute call is free.