Verifiable document-based RAG

How do you build an augmented document retrieval system that answers using relevant sources and knows when to abstain?

1960 € per session, excl. VAT

In short

This 7-hour day, excluding breaks, enables you to build a document-based RAG prototype, from preparing the corpus to evaluating source-backed answers. You compare retrieval settings, document errors and prepare a move to a pilot that incorporates access rights, document updates and quality monitoring.

Why this training?

A fluent answer can conceal poor document retrieval: an incomplete passage, an outdated version or a document the user cannot access. RAG therefore requires an explicit chain linking the question, retrieval, the context passed to the model and the answer. Optimising this chain requires retrieval errors to be separated from generation errors. A reference question set allows settings to be compared without confusing writing quality with documentary accuracy. Running case study: The fictitious organisation « Caltrive Automatismes » wants to help its support team find maintenance procedures. Its training corpus contains manuals, corrective bulletins and obsolete instructions. The prototype must distinguish between versions, cite useful passages and decline questions for which the documents provide no answer. Workplace application: Within an authorised document scope, repeat the evaluation protocol before any pilot. Appoint a corpus owner, plan reindexing and make user access conditional on validation of access rights and quality thresholds.

Objectives of this training

  • Define the document scope and expected system behaviour
  • Prepare traceable chunks with versions and access metadata
  • Assemble a retrieval and source-backed generation pipeline
  • Compare two configurations against a reference set
  • Document the technical and quality conditions for a pilot

Who is it for and what are the prerequisites?

Roles and tasks concerned

  • Software and data engineers working on document assistants
  • Technical product managers and architects scoping a RAG pilot

Prior skills

  • Have practical experience of generative AI and understand the principles of embeddings and APIs.
  • Be able to modify a script or configure a technical pipeline and interpret its logs.

What you will work on

  1. Scope use cases and evidence

    Objective: Define what the prototype must retrieve and demonstrate. RAG architecture, the distinction between retrieval and generation, functional scope and questions outside the corpus.

  2. Prepare a traceable corpus

    Objective: Transform documents into usable chunks. Extraction, cleaning, chunking, overlap, tables, identifiers, versions and permissions.

  3. Configure retrieval

    Objective: Retrieve relevant passages before any generation. Lexical and vector indexing, hybrid search, access filters, number of passages retrieved.

  4. Assemble source-backed answers

    Objective: Limit answers to the available evidence. Context construction, citations, contradictions, abstention and malicious instructions present in documents.

  5. Measure and adjust

    Objective: Choose a setting based on verifiable results. Recall of relevant passages in the results, faithfulness to sources, citation accuracy and latency.

  6. Prepare a controlled pilot

    Objective: Define the conditions for operating the prototype. Reindexing, deletions, access control, limited logging and regression testing.

What you take away

  • Architecture dossier and corpus register
  • Configured RAG prototype with a source-backed response instruction
  • Reference set and comparative results table
  • Pilot plan with risks, responsibilities and acceptance thresholds

Method and assessment

  • Short instructional inputs followed by immediate practice on the running case study.
  • Individual work followed by work in pairs, with peer review of the outputs.
  • Facilitated group review: what works, what needs correcting and what remains a human decision.
  • Preparation: Check access to a training environment and to the models authorised by the organisation. Prepare only fictitious or anonymised documents; identify their versions and access restrictions.

Observable criteria: Each chunk retains a document reference, a version and its access metadata. ; Citations point to passages that genuinely support the claims. ; The two configurations are compared using the same questions with explicit measurements. ; Cases without evidence trigger abstention; prohibited documents remain excluded. ; The pilot plan specifies updates, monitoring and error handling..

Frequently asked questions

Does a model need to be trained?

No. The work focuses on document retrieval and the context passed to the model.

Which tools are used?

A training pipeline tailored to the context, using only tools authorised by the organisation.

Can internal documents be used?

The exercises use fictitious or anonymised cases, with no unauthorised transfer of documents.

Is the prototype ready for production?

No. Security, load, operations and business validation still need to be checked before deployment.

Last updated: