Custom RAG development that answers from your data and cites the source
RAG gives a language model access to your data. We build the full pipeline: ingestion of files and databases, retrieval, re-ranking, a cited answer and a way to measure quality.
What hurts
Your team built a LangChain prototype over a weekend, and the demo looked fine. On real documents it mixes up versions, answers from the wrong section and misses tables. Nobody can say whether a change made it better or worse.
Your documents vary: scanned PDFs, Excel price lists, wiki pages, CRM records. One chunking strategy does not fit them all. Tables fall apart and headings get lost.
Then the data changes every day. Prices are updated, and the assistant still quotes the old ones. A deleted document keeps showing up in answers because nobody cleans the index.
What we build
- Connectors to your sources: files, Google Drive, your wiki, PostgreSQL, CRM
- Parsing of PDF, OCR for scans, tables, slides and HTML
- Chunking and embeddings per document type, with metadata for filters
- Vector store in Qdrant or pgvector inside your PostgreSQL
- Hybrid search and re-ranking so the right passages reach the model
- Answer generation with citations and a refusal when no source exists
- Incremental updates and removal of stale documents from the index
- Retrieval and answer evaluation suite that runs on every change
What you get
- A RAG service with an API for your products and internal tools
- Source code, ingestion config and infrastructure as code
- Question test set with references and a quality report
- Architecture docs, index refresh runbook and guide for new sources
- Dashboard with request volume, latency and cost metrics
How we measure the result
- Share of questions where the right passage is in the top results
- Share of answers with a correct citation on the test set
- Response time and cost per query
- Time from a document change to updated answers
Systems and stack
Industries
Questions
What drives the cost of RAG development?
It starts at $7,000. The number and variety of sources, the share of scans and tables, permission requirements and the hosting target set the final price. After discovery we fix the estimate within 2-3 days, and it does not change without your consent.
Which vector database do you use?
Usually Qdrant or pgvector. If you already run PostgreSQL, pgvector is often easier to maintain. For large volumes and complex filters we pick Qdrant. We decide based on your data, not on fashion.
How do you measure quality?
Together with your team we collect real questions with reference answers and sources. We evaluate retrieval and generation separately. The suite runs on every change of model, prompt or chunking, so regressions show up at once.
Can it run without sending data to third party AI providers?
Yes. Embeddings and generation can run on our own model on vLLM or on your server. Documents then stay inside controlled infrastructure, which helps with GDPR and the EU AI Act. Banks and clinics get a dedicated server.
What do you need from us?
Read-only access to sources, examples of typical questions and a subject matter expert who can check answers. We sign an NDA and a DPA before any access.
Related services
More in this section
Describe the task. We reply within one business day, the first call is with the founder.
Describe the task. The first call is with the founder.
We reply within one business day. 30 minutes to tell you whether it is worth doing and what it costs.