Pepro/RAG document assistant

RAG for business · AI assistant on your documents

Your documents.
Your questions.
The right sources.

The procedure exists. So does the manual. You still need to find the right passage, in the right version.

Pepro designs your document assistant: your teams ask a question, find an answer and can check its sources. We connect your content, your access rules and AI in a tool designed for your work.

From the first corpus to deployment · From Paris, delivered in France and abroad.

Colleagues reviewing documents together next to a computer
01 / YOUR KNOWLEDGE BASE

Share knowledge.
Without losing the source.

ProceduresGuidesTechnical files
Verifiable answersFind the passages that support them.
A controlled scopeDefine the documents and access.
A pilot before scaling upEvaluate on your real-world questions.

Understand the basics

What is RAG?
exactly?

Imagine a colleague who opens the relevant documents before answering you. That is the principle of RAG, or retrieval-augmented generation: giving the AI passages from your documentation at the moment of the question.

So there is no need to retrain a model on your whole company to get started. Preparing documents and the quality of retrieval remain essential.

01

You ask the question

In your own words, with your product reference or the name of a procedure.

02

The tool finds the passages

In the planned corpus, filtering content according to your permissions.

03

You read and check

An answer with references to the available sources.

A displayed source does not by itself prove the answer is correct. We also test the link between the passage retrieved and what the assistant concludes from it.

See the principle in practice

A question.
An answer to check.

These three scenes show the result we aim for in a project. Questions, documents and answers are fictitious and pre-written : il ne s’agit pas d’un assistant connecté à vos données.

TEAM QUESTION

What needs preparing for a new employee’s arrival?

SAMPLE ANSWER

The procedure requires an account request and equipment allocation before arrival. The team manager approves the requested access.

View source passage [1] ↓

To check: does the answer reflect exactly what is written, without adding a step?

[1] FICTITIOUS DOCUMENT RETRIEVED

Onboarding guide · version 3

Page 4 / Preparing the arrival

Read the source passage
“Before arrival, an account request and equipment allocation are prepared. Access is subject to the team manager’s approval.”

In your project, the link can open the corresponding document or passage, depending on the formats and interface chosen.

TEAM QUESTION

Does a purchase request still go by email?

SAMPLE ANSWER

Version 2 of the procedure states that requests are made in the internal tool. It replaces version 1, which required an email.

View source passage [2] ↓

To check: is the active version identified and the old one correctly excluded?

[2] FICTITIOUS DOCUMENT RETRIEVED

Purchasing procedure · version 2

Page 1 / Request process

Read the source passage
“Version 2, replaces version 1. Purchase requests are now submitted in the internal tool, replacing the email process.”

This result requires version management and withdrawal rules suited to the corpus. The model cannot guess them.

TEAM QUESTION

When is the next seminar?

SAMPLE ANSWER

I did not find a date in the documents consulted. I cannot confirm this information from this knowledge base.

To check: does the assistant flag missing information rather than suggest a plausible date?

NO SUFFICIENT SOURCE

The right answer can be:
“I don’t know.”

Abstaining is one of the cases to test. It reduces unsupported answers without guaranteeing they disappear.

Define your test cases with Pepro ↗

Where it can help you

Spend less time searching.
Make better progress.

Start with a question that often comes up in your team. The right first use case is one whose usefulness you can judge.

SUPPORT & MAINTENANCE

Find the right instruction

A manual, a reference, a check to carry out: the assistant helps prepare an answer from authorised documents.

TECHNICAL & R&D TEAMS

Explore your files

Locate a passage in long documentation and connect information, with links back to the original texts.

INTERNAL KNOWLEDGE

Pass on what already exists

Help a new employee find the procedures, guides and contacts mentioned in your knowledge base.

Plans and technical documents open on a work table

PDFs, plans, spreadsheets: their content must be usable before it can be searched.

What makes the difference

A useful answer starts
with a well-processed document.

Are your PDFs actually readable?

A scan requires character recognition. A table or plan can lose its meaning during extraction. We review representative files before choosing the method.

Which version is the reference?

Dates, duplicates, superseded documents: update and withdrawal rules are defined with you. They prevent an old instruction from competing with the new one.

Who can see what?

Permissions must apply at retrieval, before passages are sent to the model. We plan tests between authorised and unauthorised profiles.

Your project with Pepro

Build on your use cases.
Evaluate with your teams.

You bring knowledge of your business. We handle document, software and, if needed, hardware integration. The project moves forward within an agreed scope, with deliverables you can review.

Scope a first pilot ↗
01 / SCOPING

A use case and clear criteria

Users, documents, typical questions, expected answers and situations where the assistant must abstain.

02 / BUILD

A suitable document pipeline

Extraction, classification, retrieval, model and interface. Supported connectors and formats are specified in the proposal.

03 / EVALUATION

Tests that inform the decision

We review the relevance of passages, the accuracy of answers, citations, access and response time on the chosen cases.

04 / HANDOVER

A tool with a plan for what comes next

Documentation, onboarding, corpus updates and follow-up terms: deliverables and responsibilities are defined before go-live.

Choices to make together

An architecture that respects
your constraints.

RAG describes a method, not a hosting location or a confidentiality guarantee. Pepro helps you review the full journey of the data.

Local or hosted?

We specify where documents are stored and where retrieval and the model run. Access, logs, backups and external services are part of the study.

If your project needs GPU infrastructure →

Existing tool or dedicated integration?

If your tools already cover the need, that is worth checking. A specific integration is justified by your sources, access, interface or operating requirements.

Discover our software support →

Budget and scope

A quote that explains what you are paying for.

The cost depends on formats, file condition, connectors, users and the level of control expected. Our proposal separates setup, resources and any licences, then operations and upgrades.

A limited first corpus makes it possible to price a pilot, then decide on next steps from the observed results.

Before you decide

Your questions,
straight answers.

A document need can start small. What matters is knowing what you want to find and how to judge the result.

RAG, fine-tuning or quantisation: our guide (in French) →
Does RAG eliminate hallucinations?

No. Retrieval can miss a passage, the document can be outdated and the model can misread what it is given. We plan tests on answers, their sources and unanswerable cases. For sensitive decisions, human review still needs to be organised.

Can we use our PDF, Word, Excel or SharePoint files?

These are sources to assess during scoping. Support depends on the actual format, tables, scans, permissions and connection options. We review samples and confirm the project’s formats and connectors; we do not promise identical reading of every file.

Do my documents need to be perfectly organised?

Not for a first discussion. A small representative set lets us spot duplicates, versions, extraction problems and access. The preparation needed is then priced. Loading more files does not guarantee better answers.

Do we need to buy a GPU server?

Not necessarily. Equipment depends on the model, retrieval volume, concurrent use and the chosen processing location. We review existing resources and hosting options before proposing hardware. Pepro can also supply and prepare GPU infrastructure if the need justifies it.

Are our documents used to train a model?

In a standard RAG, passages are retrieved and then given to the model to answer; this is not retraining in itself. Retention and reuse conditions do, however, depend on the services used and their configuration. We review them with you before choosing the architecture.

RAG or fine-tuning: how to choose?

RAG provides knowledge to consult. Fine-tuning adapts a model to a behaviour or task from examples. They can complement each other, but we start by identifying what your use case actually lacks. Understand the differences →

What budget and timeline should we plan?

They depend on the corpus, connectors, access, interface and expected tests. The proposal specifies the pilot’s scope, deliverables, schedule and costs, as well as any recurring expenses. A reliable estimate first requires seeing a few representative cases.

Who keeps the assistant up to date after the pilot?

We define the update frequency, document withdrawal, corpus owners and technical follow-up. Training and support can be planned. Operating terms are agreed with your team, according to the chosen scope.

The first step

What question would you like
stop searching for hours?

Describe the documents concerned, the people who use them and a recurring question. Pepro helps you define a first scope and the criteria to evaluate its value.

To prepare our discussion

01 · Your main document types

02 · Two or three real-world questions

03 · Your access and hosting constraints

A description is enough for the first contact. We will then agree on sharing suitable samples.

Illustrative photographs: Alexander Suhorucov and Yaroslav Shuraev / Pexels. They do not show a Pepro team or project.

Your choice is kept for six months in this browser. You can change it at any time via “Manage cookies” at the bottom of the site.

Loading…
Back to top