Eight Mile · AI integration · Retrieval-augmented generation

Answers from your own documents, with the source attached.

Q: Do we need to move our documents?

No. The pipeline reads them where they are — a drive, a wiki, a ticketing system, a database — and keeps its own index in step. The originals stay the record.

Q: OpenSearch or pgvector?

pgvector when you already run PostgreSQL and the documents sit comfortably beside your data: one database is one less thing to operate. OpenSearch when you need strong keyword search alongside vectors, or the collection is large. We say which, and why, after seeing the data.

Q: How do you stop it showing one team another team’s documents?

Permissions are attached to every passage when it is indexed and filtered at query time with the asker’s identity. The model never sees a passage the person could not open.

Q: How do we know the answers are right?

By testing them. We build a set of questions with known answers from your own documents and measure retrieval and answer quality against it, before launch and on every change after.

Retrieval-augmented generation puts your policies, manuals, contracts and tickets in front of a model at the moment it answers, so it quotes them instead of the open internet. We build the pipeline end to end — ingestion, chunking, indexing, permission-aware retrieval — and the part that makes it trustworthy: citations, and a model that says it does not know.

Talk to an engineer What we build

What we build

Retrieval-augmented generation, built as software.

Ingestion

PDFs, Word files, wiki pages, tickets and database rows pulled in on a schedule or on change, cleaned, and kept in step with the source when a document is edited or deleted.

Chunking that respects structure

Documents split along headings, clauses and tables rather than every few hundred characters, so a retrieved passage still makes sense on its own.

The index

OpenSearch for keyword and vector search together, or pgvector inside the PostgreSQL you already run. Chosen for your data and the team that will operate it.

Permission-aware retrieval

A person only retrieves what they could already open. Access rules travel with each passage and are applied at query time, not after the answer is written.

How it works

The same steps, in the order they run.

Ingest

Sources are read, cleaned and split into passages, each carrying its document, its section and who may see it.

Index

Each passage is embedded and stored beside its keywords, so a search matches an exact product code and a paraphrase alike.

Retrieve

The question is searched with the asker’s permissions applied, and the best passages are ranked and trimmed to what the model needs.

Answer

The model answers from those passages only, citing each one. If none of them answers the question, it says so.

Controls

What makes it safe to ship.

Decided before the first prompt is written, and tested like the rest of the code.

Citations on every answer

Each claim links back to the passage it came from, so a reader can check it in one click and an auditor can check it later.

“I don’t know” is a valid answer

When retrieval finds nothing relevant the model is told to say so. A confident answer with no source behind it counts as a failure in the evaluations.

Evaluated on your questions

A set of real questions with known answers runs on every change to chunking, retrieval or prompt, so tuning one thing cannot quietly break another.

Repeat questions cached

An exact repeat is answered from a hash lookup, and a reworded one from a similarity search, before a model is paid to answer again.

Proof

Where you can see it working.

Our own products, each with a page of its own that shows how it is built.

Conversation infrastructure

ChatAgent

Each specialist retrieves from its own documents in OpenSearch, so a billing answer is grounded in billing policy and nothing else.

See how it is built

Model efficiency

llm-cache

Exact match first, then cosine similarity over an HNSW index in pgvector — the same Postgres extension a retrieval index can live in.

See how it is built

Straight answers

The questions people ask about it.

Do we need to move our documents?

OpenSearch or pgvector?

How do you stop it showing one team another team’s documents?

How do we know the answers are right?

The rest of our AI work

AI agents and assistants LLMOps MCP integration AI integration overview

Next step

Tell us which documents people keep asking about.

Say where the documents live, roughly how many there are and who asks the questions. We will tell you what an answer engine over them would take.

Meet the engineers who would do the work

Talk to an engineer [email protected]