Semantic search over your documents, with verifiable origin.

When a company wants an AI to answer questions using its own documents (contracts, manuals, records), it normally needs to assemble and maintain several connected systems. We solve it in one place, and can show exactly which document each answer came from.

Let’s talk about your documents

Against the usual way of building this, here's the difference:

Usual stackWith Gorigami
Storing vectorized documentsA separate service (Pinecone, Weaviate) synced by hand with your databaseA Vector column in the same table as your documents
Filtering by department, date, or typePost-filtering after fetching results from the vector storeFilter and semantic search in a single indexed query
Auditing where an answer came fromBuilt by hand in the application, if it’s built at allNative EXPLAIN LINEAGE, persists across restarts
Where it runsA third party’s cloudCan run embedded, never leaving your infrastructure

Under the hood, the mechanics work like this:

How we implement it

We create a vector index over the embeddings column (CREATE VECTOR INDEX), and combine semantic retrieval with relational filters in a single query (SEARCH ... FILTER <condition>). When you need to show where an answer came from, EXPLAIN LINEAGE <dataset> traces it back to the exact document and passage. The language model that writes the final answer (Claude, GPT, or whatever you already use) is still yours: we replace the retrieval layer, not the generator.

Semantic search and auditable RAG

Already in use at:

Industries already using this

  • Legal: precedents, contract clauses
  • Healthcare: clinical records
  • Customer support: internal knowledge base
  • Government & compliance: where auditing is mandatory
  • Internal knowledge management at any mid-size company

Before we talk, one clarification:

What we don’t solve

We don’t generate text or handle chunking/embedding of your documents: that still needs an embedding model (OpenAI, Cohere, or a local one) and a chunking library (LangChain, LlamaIndex, or your own). We also don’t replace the model that writes the final answer. We solve the retrieval layer, not the full RAG pipeline.

Frequently asked questions

Do you replace ChatGPT or Claude?

No. You still choose the model that generates the final answer; we make sure it finds the right information and can prove where it came from.

Do my documents leave my company?

Not if you don’t want them to: the engine can run embedded inside your own infrastructure, without sending documents to a third party.

Does it work with non-English documents?

Yes. Search quality depends on the embedding model used, not on the database.

How long does it take to implement?

It depends on the volume and format of your documents. A 20-minute diagnostic gives us a real estimate.

Chat with us