From sources to passages

How can you transform a heterogeneous document collection into a traceable, evaluated corpus that can be used by a RAG?

1960 € per session, excl. VAT

In short

This 7-hour day, excluding breaks, enables participants to prepare a document corpus for retrieval-augmented generation, or RAG. Participants select sources, clean texts, structure passages and their metadata, and then measure their retrievability in a provided environment, without designing or deploying the infrastructure.

Why this training?

A retrieval engine may retrieve an old instruction rather than a correction if versions, dates and relationships between documents remain implicit. Document quality therefore determines the relevance of the passages passed to the model. Cleaning alone is not enough: splitting a condition from its exception can make an answer misleading. The training links each transformation to a check using work-related questions and a verifiable record. Running case study: The fictitious organisation Avelune Équipements wants to provide support for its technicians using manuals, procedures and correction bulletins. Its collection contains competing versions, compatibility tables and documents restricted to certain departments. The exercises use only this fictitious case study or anonymised data. Transfer to the workplace: Apply the preparation framework to an authorised document set, have the source owner validate the permissions and hand the corpus over to the RAG lead, together with its evaluation results and remaining anomalies.

Objectives of this training

  • Select sources according to their authority, currency and usage rights.
  • Clean extracted texts without removing domain-specific information.
  • Split documents into passages that can be understood outside their immediate context.
  • Assign provenance, version and access metadata.
  • Measure the retrieval of expected passages and document corrections.

Who is it for and what are the prerequisites?

Roles and tasks concerned

  • Information specialists and knowledge base managers supplying content to a domain-specific assistant.
  • Quality leads, technical support staff and content managers involved in a RAG project.

Prior skills

  • Understand the principle of RAG and distinguish document retrieval from generation.
  • Use a generative AI assistant and work with simple structured files.
  • Know the documentation rules and access restrictions relevant to your work.

What you will work on

  1. Define the relevant document collection

    Objective: Link sources to technicians’ needs. Document authority, domain coverage, obsolescence, confidentiality and the distinction between corpus and model.

  2. Extract and clean

    Objective: Preserve meaning during normalisation. Optical character recognition errors, repeated headers, tables, encoding and deduplication.

  3. Define the passages

    Objective: Choose a segmentation approach suited to technical instructions. Heading-based structure, segment size, overlap and keeping exceptions with their rules.

  4. Record provenance and access

    Objective: Make every passage traceable and filterable. Stable identifiers, location, dates, versions, access scopes; a metadata field alone does not enforce any permissions.

  5. Evaluate retrieval

    Objective: Measure the presence of expected evidence. Annotating relevant passages, ranking results and separating retrieval errors from generation errors.

  6. Correct and hand over the corpus

    Objective: Deliver a reusable and maintainable set. Targeted corrections, comparison with parameters held constant, updating and removal of obsolete versions.

What you take away

  • Source inventory and inclusion decisions.
  • Cleaned and segmented corpus with metadata.
  • Transformation and anomaly log.
  • Annotated question set and comparative retrieval report.

Method and assessment

  • Short instructional sessions followed by immediate practice on the running case study.
  • Individual work followed by work in pairs, with peer review of the outputs.
  • Guided group discussion: what works, what needs correcting and what remains a human decision.
  • Preparation: Identify three types of work-related documents and their usual defects, without sharing confidential content. Check access to the tools authorised by the organisation and to the provided evaluation environment.

Observable criteria: Every passage retains a source identifier, a version and a verifiable location. ; A decision is recorded for duplicates and obsolete documents. ; Checked passages preserve conditions, exceptions and units from tables. ; Access restrictions are recorded and the expected controls are made explicit. ; The percentage of questions with a relevant passage among the top five results is calculated before and after correction..

Frequently asked questions

Are programming skills required?

No, the practical work uses a provided environment; an understanding of structured fields is required.

Can internal documents be used?

Only if they are anonymised and authorised; the fictitious case study remains available for all exercises.

Can AI rewrite the procedures?

Not freely: every transformation must preserve meaning and be checked against the source.

Will the corpus be deployed?

No. The day produces an evaluated corpus to be handed over to the team responsible for integration.

Last updated: