AI RAG Pipelines
Answers grounded in your own documents, with citations, an abstention when the corpus can’t answer, and an eval score behind every change.
Book an eval ↗
What is a RAG pipeline?
A RAG pipeline — retrieval-augmented generation — answers a question by finding the right passages in your own documents first, then asking a model to write an answer using only those passages. The model supplies the language; your content supplies the facts.
Almost every bad RAG answer is a retrieval failure rather than a model failure: the passage was never indexed, the chunking split a table in half, or the query didn’t share any words with the answer. So we build the graded question set before the pipeline, treat recall as the number to move, and make the system say it can’t find something instead of improvising.
What a RAG pipeline includes
A graded eval set
At least 150 real questions with agreed answers, written with your subject experts. It’s the only way to know whether a change helped or just felt better.
Honest document ingest
Parsing that survives scanned PDFs, tables and appendices, with chunk boundaries chosen by document structure rather than character count.
Hybrid retrieval
Keyword and vector search together, then a reranker — because part numbers, clause references and codes break pure embeddings.
Citations by default
Every answer points back to the passage it came from, so a reader can check it in one click instead of trusting it.
Permission-aware search
Retrieval respects who is asking, so the pipeline can’t summarise a document the person was never allowed to open.
Freshness and reindexing
Scheduled and event-driven updates with a visible index age, so nobody gets answered from last quarter’s policy.
How we build a RAG pipeline
Collect the questions
We gather real questions from the people who ask them today and agree what a good answer looks like. That set becomes the acceptance test.
Get the corpus in properly
Parsing, cleaning and chunking, plus a report on which documents are unusable as they stand. Sometimes the fix is the document, not the pipeline.
Retrieval bake-off
We score several retrieval configurations against the eval set and choose on measured recall, not on which one demoed nicely.
Wire the answer layer
Generation with citations, refusal behaviour and a tone your users recognise, behind a prompt that lives in version control.
Score it in production
Retrieval traces, feedback capture and the eval suite rerun on every change, so quality drift shows up as a chart rather than a complaint.
AI RAG Pipelines FAQ
A scoped pipeline over a defined corpus typically runs $18k to $50k, plus model and hosting costs that usually sit in the low hundreds per month at internal-team volumes. Building the eval set is a line item, not an afterthought.
The tools we build with
Retrieval decides the quality, so that’s where the budget goes — and the model itself stays a component we can replace on a Tuesday.
Retrieval & index
Qdrant
pgvector
ElasticsearchPostgreSQLModels
Anthropic
OpenAI
Hugging Face
Mistral
OllamaPipeline
Python
LangChain
FastAPI
AirflowEvals & ops
Weights & Biases
OpenTelemetry
Grafana
DockerRelated work
Related reading
Have documents nobody can get answers out of?
Send us the corpus and a dozen questions people genuinely ask. We’ll come back with an eval plan, a retrieval approach and a scoped first pipeline.