RAG development services

How it works

How a RAG system gets from your documents to a cited answer

Retrieval-augmented generation (RAG) means the system first finds the passages that answer a question and then has a language model write the answer from them, with links back to the source. Most of the work, and most of the failures, happen before the model writes a word. Plan the pipeline in seven steps.

1. Documents. Decide which sources the answers may come from, who owns them and how they are updated.
2. Parsing and OCR. Turn text, tables and scans into clean pieces (chunks) that can be cited.
3. Vector and keyword index. Store each chunk as a semantic vector and as words for keyword search.
4. Hybrid search. Combine semantic and keyword results for every question.
5. Ranking. Score the candidates against the question and keep only the few that answer it.
6. Answer with citations. The model answers only from those passages and links each claim to its source.
7. Checks on a golden set. A fixed list of real questions with known answers runs before every release.

Documents. Start with the questions, not the files. Collect a few dozen questions your users really ask, then find the documents that answer them. Sources nobody maintains give answers nobody trusts, so name an owner for each one.

Parsing and OCR. Many wrong answers start here. A table flattened into one line, or a scanned page with no text layer, is invisible to search. Split documents along their headings rather than at a fixed length, and keep the document title, the heading and the page number with each chunk, so the answer can cite them.

Vector and keyword index. Semantic vectors find passages that say the same thing in other words. Keyword search with BM25 catches exact terms such as product codes, names and clause numbers, which a semantic match can rank low. We work with Qdrant and Pinecone as vector databases, and with PostgreSQL when the data already lives there.

Hybrid search and ranking. Run both searches for every question and merge the results. Then a reranking model scores each candidate passage against the question, and only the best few go to the language model. More context is not better. Passages that do not answer the question give the model something to get wrong. When a question needs several lookups, agentic retrieval lets the model search again with a sharper query before it answers.

Answer with citations. Ask the model to answer only from the passages it was given, to cite each one and to say so when they do not hold the answer. A link to the source lets a user check an answer in seconds, and an honest "this is not in the documents" is worth more than a confident guess. Nestwise gives its parenting advice with citations to trusted sources (Nestwise case study). Why a confident answer can still be wrong is the subject of our post on the Dr. Fox effect and ChatGPT.

Checks on a golden set. Keep the questions from the first step, each with its expected answer and the source it should cite. That list is your golden set. Score retrieval and answers separately. Check whether the right source came back, and whether the answer was right given that source. Then you know whether a failure starts in the index or in the prompt. Run the golden set after every change to the documents, the index, the prompt or the model, and compare each run with the last one.

A chatbot that answers from your documents

A chatbot like this is the most common RAG project, and the pipeline above is its core. Around it sit the parts users notice. The conversation keeps its context, each user gets answers only from documents they are allowed to open, and new documents can be added without a developer. Nestwise shows the pattern. Its RAG-based engine recalls family context across sessions and turns a parenting question into a step-by-step action plan. We delivered it as a working proof of concept, which let the client validate product-market fit and collect real user feedback (Nestwise case study).

LLM document analysis when a mistake is costly

When a language model reads contracts, reports or resumes, the job changes. Instead of answering a question, the system pulls fields and findings out of each document. Two habits keep it safe. Make the model point to the place in the document behind every value it extracts, so a person can check it in one click. And send unclear or high-stakes results to a person before they feed a decision. Persona, our own platform, uses AI to parse resumes, match candidates and validate company data (Persona case study).

When RAG needs a knowledge graph

A vector index finds passages that look like the question. It does not know that a product belongs to a product line, or that one contract amends another. A knowledge graph stores such links as entities and relations, so retrieval can follow them. Add one when the right answer depends on connections between records rather than on a single passage, and keep it next to the index, not in place of it. Start without a graph, and add it when the golden set shows questions that plain retrieval keeps missing. We build both, the index and the graph, and connect them in one retrieval step.

Our other AI work, from computer vision to live captions, is on the AI case studies page, and the wider service is AI transformation.

Related reading

• 

  What is the Dr. Fox effect, and why a confident ChatGPT answer can still be wrong

  A fake lecture in 1970 shows why a confident tone is not proof. How ChatGPT's convincing but sometimes wrong answers affect developer communities, and how to ask for evidence, not confidence.

• 

  How We Build AI Systems That Think, Act, and Orchestrate

  A deep dive into the AI systems we build across workflow automation, content intelligence, knowledge copilots, and fintech document analysis, showcasing how agentic orchestration and RAG deliver real operational impact and measurable ROI.