Skip to content

Offer 03 · RAG and knowledge systems

RAG systems that answer from your documents and show their sources

Retrieval built on your own data, with permissions enforced on every question, and an evaluation run before each release.

Scope my knowledge systemSee the types we build

Types we build

Six kinds of knowledge system, each built around one job

  • 01

    Internal knowledge assistant

    Staff ask questions in plain words and get answers with their sources.

    1. Staff question
    2. Search permitted sources
    3. Answer with citations

    Example workflows

    • An assistant that answers HR and benefits questions from approved policy documents.
    • A sales helper that finds the current pricing terms and the clause each answer relies on.
    Build this agent
  • 02

    Customer answer assistant

    Answers for customers, drawn only from published, approved material.

    1. Customer asks
    2. Retrieve published content
    3. Answer or hand off

    Example workflows

    • A support assistant that answers from the help centre and hands billing disputes to a person.
    • A product guide that cites the manual page behind each answer.
    Build this agent
  • 03

    Permission-aware search

    Search across document stores, with the same access rules as the source.

    1. Person searches
    2. Filter by their access
    3. Return only what they may read

    Example workflows

    • Contract search where each user sees only the matters they are on.
    • Engineering docs search that respects repository and team access.
    Build this agent
  • 04

    Data pipeline for AI

    Clean, current data that retrieval can rely on, kept up to date.

    1. Source changes
    2. Clean and index
    3. Stale copy removed

    Example workflows

    • A nightly pipeline that updates the policy index when a source document changes.
    • A deduplication step that keeps one current version of each procedure.
    Build this agent
  • 05

    Compliance answer desk

    Answers compliance questions from approved policies, with an audit trail of every answer.

    1. Question arrives
    2. Retrieve approved policy
    3. Answer logged for audit

    Example workflows

    • A desk that answers policy questions from approved documents and records each source used.
    • A check that confirms an answer cites the current version before it is sent.
    Build this agent
  • 06

    Document review

    Finds clauses and dates across contracts, with each finding linked to its page and logged.

    1. Document uploaded
    2. Clauses extracted
    3. Reviewer signs off

    Example workflows

    • A contract review that lists renewal dates and links each one to its clause.
    • A review queue where each change to a finding is recorded with who made it.
    Build this agent

Where these systems fail

Four ways knowledge assistants go wrong

  • 01

    Nobody can verify the answers

    Agent does
    Every answer is built from retrieved passages and shows the source it came from, so a reader can check it.
    Person approves
    Answers that will be used in customer-facing or regulated work.
    Scope this situation for my business

    A model answering from memory sounds confident whether or not it is right, and nobody can tell which.

    • Answer given with no source attached
    • Staff spend an hour checking whether it is true
    • Wrong answer repeated in a customer email
    BeforeAfter

  • 02

    The assistant shows people documents they should not see

    Agent does
    Permissions are copied from your source systems at index time, and retrieval checks them on every question.
    Person approves
    Changes to who can access which source.
    Scope this situation for my business

    Search that ignores permissions puts restricted files into answers for people who were never meant to read them.

    • Restricted contract text appeared in a general answer
    • No record of which documents were used
    • Access was only checked after the answer was shown
    BeforeAfter

  • 03

    Data is scattered, and some of it is stale

    Agent does
    Sources are cleaned and deduplicated during ingestion, and each document keeps its version and last-updated date.
    Person approves
    Which sources count as the current version.
    Scope this situation for my business

    The same policy exists in three folders and two versions, so the answer depends on which copy the search finds first.

    • Three folders, two versions of one policy
    • Answer changed depending on which copy was found
    • Nobody sure which version is current
    BeforeAfter

  • 04

    Retrieval works on ten documents and fails on ten thousand

    Agent does
    Every release runs a fixed set of questions with known answers, so a drop in quality shows before users see it.
    Person approves
    Any release that changes the question set or the thresholds.
    Scope this situation for my business

    A setup that looks right on a small sample loses the correct passage once the corpus grows, and nobody notices until users complain.

    • Demo on a small sample looked correct
    • Correct passage missed once the corpus grew
    • Users reported it before anyone measured it
    BeforeAfter

Questions buyers ask

Before you scope a project

What is RAG?
Retrieval-augmented generation means the model answers using passages retrieved from your own documents, rather than from what it learned in training. The retrieved passages are shown with the answer, so the source can be checked.
When is RAG the wrong choice?
When the answer does not depend on your documents, or when the documents are too few to need search. A plain prompt or a simple lookup can be better. We check this in the scoping phase before building anything.
How does the system respect document permissions?
Each document keeps the access rules from its source system. Retrieval filters passages by the person asking, so a user only receives sources they could already open. Access changes in the source flow through on the next index update.
How do you know the answers are good?
We build a fixed set of questions with known answers from your own material. Every release runs the set, and it measures whether answers are supported by a source and whether the right document was found.
Which model and vector database do you use?
It depends on your data, your hosting rules and your cost limits. We compare options during scoping and record the choice, so you can change it later without rebuilding the pipeline.
Can the system run inside our own environment?
Yes, when your rules require it. Documents, indexes and logs can stay in infrastructure you control. The deployment choice is part of the scoped phase and is approved before the build starts.
What does the audit trail record?
Every action the system takes: the input, the source or factors used, the output, and who approved it if a person was involved. The trail is kept in storage you control, so your team can answer an audit question from the record.

Case studies

Writing

Related offers

Other ways we build AI

Start with one source

Tell us the documents your team cannot find

We scope the first knowledge source with a fixed price for each step, before any build starts.

Scope my knowledge system
  1. 01

    Share the source

    Tell us where the documents live and who should read them.

  2. 02

    We scope the phase

    A fixed-price plan for the first source, with the approval step named.

  3. 03

    We build in phases

    You review retrieval quality before the next source is added.Every change is logged in the audit trail.