We build retrieval-augmented generation systems that answer from your documents, databases and knowledge bases, cite every answer back to its source, and are measured against real questions before they reach your users.

Get a production-readiness review How we build it

01 · What we build

Ingestion & chunking

Connectors for the places your knowledge already lives: documents, wikis, tickets, databases and catalogues. We clean it and chunk along the structure each document already has, with its title and headings attached to every chunk.

02 · What we build

Hybrid retrieval & re-ranking

Vector search for meaning, a keyword index for exact terms like product codes and names, and a re-ranker on top. Metadata filters and permission checks run before anything reaches the model.

03 · What we build

Citations & guardrails

Every answer points to the passages it came from. When the sources do not support an answer, the system says so instead of guessing, and sensitive data stays out of prompts and logs.

04 · What we build

Evaluation

A test set of real questions with known good answers, scored for retrieval and for answer quality on every change. A change that makes things worse is caught before release, not after.

05 · What we build

Freshness & operations

The index is a pipeline, not a one-time load. Sources are re-indexed as they change, every chunk carries a date, and latency, cost and answer quality are monitored in production.

02 · How we build it

Retrieval first.

Questions before vectors
We start by collecting real questions and the answers a domain expert would give. That set decides every later choice, from chunking to model.
Retrieval measured on its own
If the right passage never reaches the model, no model can fix it. We measure and fix retrieval before touching prompts or swapping models.
Cite or decline
An answer has to point to its source. When the sources are silent, the honest answer is that the system does not know.
Built to stay fresh
Re-indexing, freshness dates and monitoring ship with the first release, because a RAG system that is never updated quietly goes stale.

03 · Is RAG the right tool?

Sometimes it is not.

Choose RAG when

Answers must come from documents that change
Users need to see the source
Access differs by user or team

Choose fine-tuning when

You need a consistent style or format
The task is narrow and stable
A smaller, faster model has to do the job

Choose neither when

Search or simple rules would do
There is no reliable source data yet
Nobody will act on the answer

RAG vs fine-tuning, in depth

04 · Tech stack

Chosen per project.

Vector stores

pgvector
Qdrant
Weaviate
Pinecone
Elasticsearch / OpenSearch

Retrieval

Hybrid keyword + vector
Cohere Rerank
Cross-encoder re-rankers
Metadata & permission filters

Frameworks

LlamaIndex
LangChain
Custom pipelines

Evals

RAGAS
Deepeval
Custom harnesses

05 · Common questions

Before you brief us.

What data can a RAG system use?: Most text your team already relies on: documents and PDFs, wikis, policies, support tickets, product catalogues and database records. We start with an audit of which sources are reliable and current, because the system can only be as good as what it retrieves.
Will it make things up?: No system can promise it will never be wrong. We design so answers come only from retrieved sources, show their citations, and decline when the sources do not support an answer, and we measure that on a test set before launch and after every change.
Can it respect who is allowed to see what?: Yes. Retrieval filters by each user’s permissions before anything reaches the model, so people only get answers built from documents they are allowed to read.
Do we also need to fine-tune a model?: Usually not at first. RAG handles knowledge that changes; fine-tuning handles style, format and narrow, stable tasks. Some systems use both, and we decide with your data rather than by default.
Can it run in our own cloud?: Yes. We deploy into your cloud environment and work with hosted models from OpenAI, Anthropic and Google, or with open models you run yourself.

Available for new engagements

Put your knowledge to work.

30 minutes, one of our seniors, no slide deck. By the end of the call you will know whether we are the right team, and if not, who is.

Get a production-readiness review Email info@octalcode.com

Or call +1 (512) 710-5701

Let's scope it.Instant answers · free project scoping