Enterprise RAG and document intelligence
The short answer
Enterprise retrieval-augmented generation (RAG) answers questions and extracts facts from your own documents and records, with a citation back to the source for every claim. Branemind builds these systems with parsing and normalisation before retrieval, hybrid search with reranking, access control at query time, and evaluation on your own questions.
Retrieval is easy to demo and hard to trust
Pointing a model at a folder of PDFs produces a convincing demo in an afternoon. It stops being convincing when an answer has to be defended: when the source is a scanned statement with a broken table, when two documents disagree, when the user should not be allowed to see half the corpus, or when nobody can tell whether a retrieval change helped. In our underwriting work, most of what first looked like the model hallucinating turned out to be parsing bugs upstream of it.
What it costs today
- Answers without a source a reviewer can open, so every answer is re-checked by hand
- Exact identifiers such as part numbers and clause references missed by embeddings alone
- Documents surfaced to users who should never have seen them
- No evaluation set, so quality is argued about rather than measured
What we build, and where we stop
The second list matters as much as the first. A boundary that is agreed in writing before the build is the difference between a system your compliance team signs off and one they discover.
What we build
- Layout-aware parsing and normalisation of PDFs, scans, tables and statements into typed records
- Hybrid retrieval: keyword search for exact identifiers, dense retrieval for meaning, a reranker to settle between them
- Span-level citations, so every extracted field or answer links to the page and line it came from
- Permission-aware retrieval that filters by the user's access before anything reaches the model
- An evaluation set built from your real questions, with retrieval and answer quality measured separately
- Structured knowledge, such as compatibility or supersession, modelled as a graph where a prompt would guess
What we do not do
- Index a corpus without agreeing who may see which documents
- Report a confidence score that has not been calibrated against reviewed answers
- Promise zero human review on high-stakes documents. The target is a review queue that shrinks with a reason for everything still in it
- Build a chat interface where a search box or a report would serve users better
How it works
- 01
Collect the real questions
We start from the questions people actually ask and the documents that should answer them, including the ones that currently go unanswered. These become the evaluation set before any retrieval is built.
- 02
Parse and normalise
Documents are converted into structured records that keep tables, clause boundaries, defined terms and page references intact. Nothing reaches the model as an unstructured dump of PDF text.
- 03
Retrieve hybrid, rerank
Keyword and dense retrieval run together, and a cross-encoder reranker orders the candidates. Retrieval quality is measured on its own, before any generation is involved.
- 04
Answer with citations
Answers and extracted fields carry the span they came from. Anything below the agreed confidence threshold routes to a reviewer instead of being asserted.
- 05
Close the loop
Reviewer corrections feed back into the evaluation set, so the threshold and the retrieval configuration improve on evidence rather than on opinion.
What it connects to
- Document stores and shared drives
- Postgres with pgvector, and other vector or search infrastructure
- Line-of-business systems and product catalogues
- Identity providers, for permission-aware retrieval
- Model providers, with the choice kept swappable
- Deployment into your own cloud account where data residency requires it
What you need in place
We would rather tell you this before a proposal than during one.
- A representative sample of the documents, including the messy ones
- A list of real questions with known-good answers, or someone who can write them
- A clear access model: who may see which documents
- A decision on where documents may be processed, made before build
Controls, approvals and what happens when it fails
Citation or no answer
Every answer and extracted field links to its source span. An answer that cannot be grounded is not returned as fact.
Access control at retrieval
Permissions are enforced when candidates are retrieved, not after generation, so restricted text never enters the prompt.
Calibrated thresholds
Confidence thresholds are set against reviewed answers. Below the threshold, the item goes to a person with the candidates attached.
Retrieval is versioned
Parsers, chunking, retrieval and reranking configuration are versioned like code and measured against the evaluation set before release.
The proof
Search that understands a part, not just its name
Forty thousand SKUs, half of them near-identical, and buyers who search by symptom rather than part number. A reranker and a graph fixed what keyword search never could.
- Search-to-cart
- +41%
- Zero-result queries
- −78%
- Support tickets
- −1/3
- 'which part do I need'
What we measure, and what we do not promise
- What we measure
- Retrieval recall on your evaluation set, answer correctness, citation coverage, zero-result and escalation rates, reviewer time per document, latency and cost per query.
- What a measurement looked like
- For a dental equipment manufacturer's catalogue of around forty thousand SKUs, hybrid retrieval with a reranker and a product graph cut zero-result queries by 78% in one engagement, measured against the search baseline before the work began.
- When we say no
- If the answers live in one structured database, a well-designed query or report beats retrieval. If the corpus is small and stable, a curated FAQ may be enough.
What this ships on
Questions we get asked
What is enterprise RAG?
Retrieval-augmented generation retrieves relevant passages from your own documents or data at question time and gives them to a language model to answer from, instead of relying on what the model learned in training. In an enterprise setting it also has to respect who may see which documents, cite its sources and be measured on your own questions.
How does RAG keep private documents private?
By filtering on the user's permissions at retrieval time, so restricted text never reaches the prompt, and by deploying the retrieval index and runtime where your data residency rules require, including your own cloud account. Model provider choice can be constrained for the same reason.
How do you measure RAG accuracy?
Retrieval and generation are measured separately. Retrieval is scored on whether the right passages come back for a set of real questions. Answers are scored for correctness and for whether every claim is supported by a cited passage. Without that split, a retrieval failure looks like a model failure.
Why use hybrid search instead of embeddings alone?
Embeddings are good at meaning and poor at exact identifiers. Part numbers, clause references, invoice IDs and names are matched far more reliably by keyword search. Running both and reranking the combined candidates handles the person who knows the exact term and the person who describes the problem.
Should we fine-tune a model or use RAG?
For answering from documents that change, RAG is usually the right starting point: the knowledge stays in your documents, updates without retraining, and every answer can be cited. Fine-tuning is better suited to changing a model's format or behaviour than to teaching it facts.
Assess your document workflow
A readiness read on data, access and evaluation before build.