01
New projects
A paragraph about the problem is enough to start. We reply within two working days, usually with questions rather than a quote.
AI · Service RAG development: retrieval over your own data
Retrieval-augmented generation that respects source permissions, cites what it used, and says so instead of guessing when the index has no answer.
In short
Retrieval-augmented generation (RAG) lets a language model answer from your own documents: the system finds the relevant passages first, then gives them to the model along with the question. We build retrieval systems that respect who is allowed to see each document, cite the passages they used, and say "not found" instead of inventing an answer.
- How to engage
- Product build
- Practice
- AI
Signs this is the right service
- Staff or customers ask questions whose answers live in documents, tickets or a knowledge base.
- A chatbot answers confidently and is sometimes wrong, with no way to see where the answer came from.
- Documents have different access rules and the assistant must not leak across them.
- Search returns the right document on page three.
What you get
Ingestion pipeline
Parsing, chunking and embedding for your actual formats — PDFs, tickets, wikis, tables — with re-indexing when a source changes.
Permission-aware retrieval
The access rules of the source system are applied at query time, per tenant and per user, and tested by trying to cross them.
Hybrid search and reranking
Keyword and vector search combined, then reranked, because either one alone misses exact terms or misses meaning.
Citations and refusals
Every answer links to the passages it used. When retrieval finds nothing relevant, the system says so.
Retrieval evaluation
A question set with known source passages, so retrieval quality is a measured number that is checked on every change.
The process
- 01
Audit the sources
What exists, who may read it, how often it changes, and which questions people actually ask.
- 02
Build the question set
Real questions paired with the passages that answer them, agreed with the people who own the content.
- 03
Tune retrieval before generation
We get the right passages retrieved first. A better model cannot fix the wrong context.
- 04
Ship with monitoring
Unanswered questions and low-confidence answers are logged, which shows which documents are missing.
pgvectorPostgresHybrid searchRerankersClaudeTypeScriptPython
RAG & retrieval systems: common questions
Do we need a separate vector database?
- Often not. If your data is already in Postgres, pgvector keeps retrieval next to the rows and their permissions. We move to a dedicated vector store only when the scale or the query pattern needs it.
How do you reduce wrong answers?
- Three ways: retrieve better passages, make the model answer only from what was retrieved, and return "not found" when nothing relevant comes back. Each is measured with the evaluation set rather than judged by eye.
Will our documents be used to train a model?
- We use provider APIs under terms that exclude training on your data, and we tell you exactly where your data is sent and stored before anything leaves your systems.
Can it work with documents that change every day?
- Yes. The ingestion pipeline re-indexes a source when it changes and removes passages when a document is deleted or its permissions change.
06 / Let’s build
Bring us the difficult problem.
02
Careers
Engineers, designers and researchers who like systems that are hard to get right. Send something you built and what you would change about it now.
03
Studio
Based in India and working remotely, with overlap hours agreed up front so your team always has a time to reach us.
04
Read
Guides on AI agents, RAG, MCP, tokenization and chain abstraction, written by the engineers who build them.