What is a production RAG system?
Retrieval-augmented generation looks up passages from a corpus you own, then writes only from those passages. Production RAG adds chunking that matches how operators search, hybrid retrieval, source cards, and a filter that refuses to speak when evidence is missing. It is a knowledge product, not a wrapper around a public model.
Proof from private corpora
The private RAG product recommender ranks catalog items from the merchant’s own product knowledge plus live session behavior, with explanations merchandisers can read. ChainTech AI Chat answers corridor, FX, and compliance questions with source documents beside the reply. Audit Genie grounds PBC line items in a standards library and drops unverified rows into a fallback loop.
How we stand up retrieval
- Corpus and access. Inventory which documents, SKUs, or workpapers the system may see, who owns them, and what must never leave the tenant.
- Chunk, embed, retrieve. Design chunking and hybrid search for the actual questions (SKU attributes, corridor rules, PBC standards)—not a generic 512-token split.
- Grounded generation. Force answers or rankings to cite retrieved items. Add a reject path when retrieval is empty or confidence is low.
- Operator surface. Ship source cards, merchandising rationale, or an auditor checklist so humans can verify the system instead of trusting a blob of text.
What you receive
- Ingest pipeline and private vector (or hybrid) index for the client corpus
- Retrieval + generation API with citation objects, not just a string
- Grounding filter or fallback when evidence is missing
- A chat, ranking, or checklist UI your operators already understand
- Docs for re-ingest, evaluation queries, and access control
Related work and reading
Frequently asked questions
Can RAG run only on our documents?
Yes. Private RAG retrieves from the client corpus—catalog, policy PDFs, workpapers—not a shared public index. That is how the ecommerce recommender and ChainTech advisory chat stay defensible in review.
How do you stop hallucinated citations?
We retrieve first, then generate, and we reject answers that cannot be tied to a source. Audit Genie uses a strict grounding filter; ChainTech surfaces the documents next to each claim.
Is this the same as a chatbot with file upload?
No. File-upload chat is a session toy. A knowledge system has ingest, permissions, hybrid retrieval, eval queries, and an operator UI that survives the next document drop.
Which industries have you done RAG for?
Ecommerce catalog ranking, fintech payments advisory, and accounting/audit PBC workpapers. The pattern transfers; the chunking and grounding rules do not copy-paste blindly.