RAG & automation

A chatbot is easy.One you can trust is not.

Grounded retrieval over your own documents, agents that act inside your operational systems, and private deployment where the data cannot leave — with citations, evaluation, and permissions that carry through from the app.

Assess an AI workflow Try Anchor

Quick answer

AI automation only works when it is grounded — answers retrieved from approved documents with citations, permissions inherited from the host application, an evaluation harness on every change, and a refusal path for questions the sources do not support. Seypro builds these systems and runs one publicly: Anchor, a document-grounded assistant you can test on your own file.

Why pilots stall

Four ways an internal AI project dies before it reaches anyone.

None of these are model problems. They are engineering problems, and they show up in the same order every time.

It answers confidently and it is wrong

Ungrounded generation over company data produces fluent, plausible, unsourced answers. The fix is retrieval with citations and a refusal path — not a longer system prompt.

It leaks across permission boundaries

A single shared index means anyone who can ask can reach anything indexed. Access control has to be enforced at retrieval, inheriting the host application’s existing rules.

It works in the demo and not on the corpus

Ten curated PDFs behave nothing like forty thousand mixed documents with duplicates, scans, and superseded versions. Chunking, hybrid search, and reranking are where that gap closes.

Nobody can tell whether it got worse

Without an eval harness every change is a guess. With one, a regression is a failing test rather than a complaint three weeks later.

What we build

Retrieval, action, and the guardrails between them.

Wired into the data, permissions and systems you already run — not a separate product your team has to remember to open.

Grounded retrieval (RAG)

Retrieval over approved documents with hybrid search, tuned chunking and embeddings, and source citations on every answer. Built on pgvector or Pinecone depending on what the rest of your stack already runs.

Agents that act

Tool-using agents wired into operational systems — CRM, ERP, internal APIs. Multi-step reasoning with function calling, and human-in-the-loop at the steps where being wrong is expensive.

Private deployment

Open-weight models on infrastructure you control, via Ollama or vLLM, with GPU tuning and quantization. For when data residency, provider exposure, or predictable capacity is the constraint.

Permissions that carry through

Access control inherited from the host application, so retrieval cannot surface a document the requesting user was never allowed to open. This is the failure mode that ends internal pilots.

Evaluation harness

A test set with expected behaviour, run on every change. Without it you cannot tell whether a prompt edit improved the system or quietly broke a case nobody checks.

Honest refusals

A system that answers everything is worse than one that says it does not know. Grounding means the answer either traces to a source or does not get made.

Stack

Anthropic
OpenAI
Mistral AI
Ollama
vLLM
Amazon Bedrock
AWS SageMaker
pgvector
Pinecone
MLflow

Our own system, in the open

Anchor is a document-grounded assistant we built and run. Upload a file and question it — cited answers, honest refusals when the document does not support a claim, and nothing retained after the session. It is the argument on this page, working, rather than described.

Try it on your own file What Anchor is

The next question

Once it works, someone will ask you to defend it.

Automation that touches decisions runs into classification, explainability, and audit-trail obligations. Better to build the evidence layer alongside the system than to retrofit it under deadline.

AI governance & EU AI Act readiness

Questions

Straight answers.

What is RAG, in one sentence?

Retrieval-augmented generation means the model answers from documents fetched at question time rather than from memory, so answers can cite a source and can be refused when no source supports them.

Do we need to fine-tune a model?

Usually not. Fine-tuning changes style and format well and teaches facts badly. If the requirement is that the system knows your current documents, retrieval is the correct tool and stays correct when the documents change.

Can this run without sending data to a provider?

Yes. Open-weight models served on your own infrastructure through Ollama or vLLM, with GPU sizing and quantization tuned to the workload. Expect to trade some capability for that control.

How do you stop it hallucinating?

You constrain what it can answer from and you make refusal an acceptable output. Grounded retrieval with citations, an evaluation harness that tests refusal behaviour, and human-in-the-loop on the consequential steps.

One next step

What is the workflow you would automate first?

Bring the one that eats the most hours and has the clearest right answer. That is where retrieval and agents earn their keep — and where a failure is cheap to catch.

Assess an AI workflow The full AI practice