Retrieval augmented generation, or RAG, is how you make an AI answer from your documents instead of from the internet. EqualPixels builds RAG systems and internal knowledge assistants so your team can ask a question in plain language and get an answer that cites the exact policy, contract or specification it came from.
What is RAG?
RAG works in two steps. First the system searches your own content for the passages most relevant to the question. Then it passes those passages to a language model and asks it to answer using only that material. The model does not recall your data from training, it reads it at the moment of asking.
That is why RAG is the default architecture for internal assistants: it keeps answers current, keeps them attributable, and keeps your documents out of anyone’s training set.
RAG vs fine-tuning: which do you need?
| RAG | Fine-tuning | |
|---|---|---|
| Best for | Facts, policies, documents that change | Style, format, consistent task behaviour |
| Updating | Re-index, minutes | Retrain, days |
| Citations | Yes | No |
| Cost to start | Low | High |
If the question is “what does our documentation say”, the answer is RAG. Fine-tuning is for shaping how the model responds, not for teaching it facts. Most projects that ask for fine-tuning need retrieval instead.
What we build
- Internal knowledge assistants over policies, SOPs, wikis, past proposals and technical documentation, with permissions so people see only what they should
- Contract and document question answering, including clause lookup and comparison across versions
- Customer-facing support search grounded in your help centre, with citations shown to the user
- Engineering and product assistants over codebases, specifications and ticket history
- Data-aware assistants that combine document retrieval with live queries against your database
How we build a RAG system that is actually accurate
- Audit the corpus. Duplicated, contradictory and outdated documents are the number one cause of wrong answers. We find them before indexing.
- Chunk with structure, not by character count. Splitting on headings and sections preserves meaning that naive chunking destroys.
- Hybrid retrieval. Keyword and vector search together, then reranking, because pure vector search misses exact identifiers, product codes and names.
- Permission-aware indexing. Access control applied at retrieval, so an answer can never leak a document the user cannot open.
- Evaluate honestly. A fixed question set with approved answers, scored for correctness and for citation accuracy, run on every change.
- Show the source. Every answer links to the passage it came from, so people can verify rather than trust.
The stack
pgvector, Pinecone or Qdrant for vectors; hybrid search with BM25 and reranking; OpenAI, Anthropic Claude, Gemini or open-weight models depending on data residency; deployed in your cloud where regulation requires it.
Frequently asked questions
How do I connect ChatGPT to my company’s internal data?
Not by uploading everything into a chat window. You build a retrieval layer over your documents and expose it either through a private assistant or through a Model Context Protocol server that ChatGPT and Claude can call. That keeps access controlled and answers current.
What is RAG and when should I use it instead of fine-tuning?
Use RAG when the answer lives in documents that change. Use fine-tuning when you need a consistent output style or a narrow repeated task. They are complementary, and RAG is almost always the cheaper first step.
Is our data safe?
Your documents stay in your storage. We use enterprise endpoints that exclude data from training, and where regulation demands it we run open-weight models entirely inside your own infrastructure.
How accurate is it?
Accuracy is a property of your content as much as the model. We measure it against a fixed evaluation set and report the number rather than describing it as high. If accuracy is unacceptable, the fix is usually the source documents, not the model.
How long does a RAG build take?
A working internal assistant over a defined document set typically takes three to six weeks, with most of that time spent on content preparation, permissions and evaluation rather than on the retrieval code itself.
Can it work across several systems?
Yes. We routinely index SharePoint, Google Drive, Notion, Confluence, a help centre and a database together, with a single permission model across all of them.