Narrathèque
Try it free
English
FrançaisEnglishFrançais (Canada)
Under the bonnet of Narratheque

Search your sources.
Build links between them.

Vector RAG and LLM Wiki: two ways of giving context to an AI. The first finds useful passages, the second connects the ideas to one another.

The principle

Vector RAG feeds the answer
with your own knowledge.

A language model holds knowledge acquired during its training. To answer about your procedures, your projects or your clients, it also needs the information you hold.

RAG — retrieval-augmented generation — brings documentary material in at the moment of answering. The LLM Wiki adds a layer of notes written and connected beforehand.In both cases the sources remain essential: a summary or an answer does not replace the document that lets you check a fact.

Your question

“Which procedure applies to this client?”

+

Source extracts · Wiki notes

↓

From the question, the instructions and the context selected.

The model is given context; it is not retrained on every document.

On fournit du contexte au modèle ; on ne le réentraîne pas à chaque document.
The essential difference

Vector RAG or wiki: when is the knowledge prepared?

Vector RAG assembles the passages at the moment of the question. The wiki does part of the connecting as soon as the sources arrive. In both approaches, the model can still compose an answer on demand.

01 · Retrieve

Vector RAG

An index for searching by meaning.

On import

  1. Extract the content

    The text is retrieved; a media file can be transcribed and an image analysed.

  2. Split into passages

    Portions of text keep their reference to the original document.

  3. Build the vector index

    An embedding model translates the meaning of the passages into numerical representations.

At every question

  1. Find the nearest passages

    The question is compared with the entries in the index to select a context.

  2. Write from the extracts

    The LLM receives the question and the selected passages to build its answer.

What persists

The sources and their index. A summary does not automatically become lasting new knowledge.

02 · Organise

LLM Wiki

Notes that connect knowledge.

On import

  1. Read the new source

    The LLM spots the subjects, people, projects and facts worth keeping.

  2. Create or enrich the notes

    It prepares summaries and adds to the pages already devoted to those subjects.

  3. Connect and confront

    The notes point back to the sources and to each other; disagreements can be flagged.

At every question

  1. Consult the relevant notes

    The search rests on an organisation of knowledge already prepared.

  2. Summarise and check

    The model combines the useful notes and can go back to the original documents.

What persists

The sources, the notes and their relations. Updates call for someone to watch their quality and how current they are.

A concrete example

Vector RAG on three documents,
two different questions.

A made-up example: your base holds the documents of the Atlas project. The values below are there only to illustrate the two paths.

Document 01

Initial contract

Delivery planned: 30 June.

Document 02

Minutes of 12 June

A delay in approval is reported.

Document 03

Signed amendment

New deadline: 15 July.

Looking up a fact

“Which date appears in the amendment?”

  1. RAG finds the passage

    The request targets one fact in one identified document.

  2. Expected answer: 15 July

    The passage from the amendment is the evidence; the initial contract is not the reference for this question.

Understanding across documents

“How has the schedule changed?”

  1. The wiki gathers the history

    The Atlas note can connect the contract, the minutes and the amendment.

  2. A timeline to check

    30 June planned → delay reported → 15 July signed. The three sources let you verify that reading.

Why combine them? The note gives the overall view; the original passage confirms the exact date and the conditions that come with it.

Inside Narratheque

Vector RAG and wiki:
two paths across one corpus.

The architecture Narratheque presents combines searching the sources with structured knowledge. The diagram describes what each part does, without fixing the details of any one technical configuration.

Your sources

Documents · Web pages · Audio · Video · Images

Preparing the content

Text extraction, transcription or OCR depending on the format

Search

The vector RAG index

Passages and references back to the sources

Structuring

Knowledge wiki

Notes, summaries and relations

Your question guides the search

A precise detail, a summary, a comparison: draw on the extracts, the notes, or both.

An answer to set against the references

The model you chose drafts the answer from the material selected.

With every new import, the index and the notes concerned have to be brought up to date. A useful analysis can then be kept, once checked.
Comparing the approaches

Vector RAG and wiki:
different strengths, limits worth knowing.

A comparison of the basic mechanisms; implementations may go further.
CritèreRAG vectorielLLM Wiki
What is keptSources, passages et vecteurs.Sources et pages structurées reliées.
Work before the questionExtraction, découpage, indexation.Lecture, synthèse, rapprochements et rédaction.
Work at the questionRecherche d’extraits puis génération.Recherche de fiches puis génération, avec retour aux sources si nécessaire.
Targeted searchAdapté à la recherche d’un passage dans un grand corpus.Utile pour retrouver le contexte déjà regroupé autour d’un sujet.
Cross-document viewÀ reconstruire à partir des passages retrouvés.En partie préparée dans les fiches et leurs relations.
UpdatesRéindexer les passages ajoutés ou modifiés.Revoir les fiches affectées, leurs liens et leurs synthèses.
Cost to watchIndexation, recherche et contexte transmis au LLM.Rédaction et entretien du wiki, puis consultation.
Main limitationUn extrait manquant ou isolé peut produire une réponse incomplète.Une synthèse erronée ou périmée peut influencer plusieurs réponses.
Which path for which need?

Vector RAG: start from the question.

Find a piece of information

A passage from a training course, a contractual condition or a technical value: start with the sources and their extracts.

Understand a subject

The history of a project, the decisions around a product or positions to compare: the notes can give a first structured view.

Check a summary

Tie the overall view back to the original passages. Check the dates, the versions and any discrepancies.

The framework matters too

Search is only one part
of your AI environment.

Choice of models

Models are there to analyse and to write. Their capabilities, their cost and where they process data all count in the configuration.

Découvrir

Preparing the sources

A good transcription or OCR makes searching easier and cuts down the errors carried into the notes.

Découvrir

Access and protection

Rights over the sources and what derives from them must stay consistent. RAG or a wiki does not replace managing permissions.

Découvrir
Frequently asked questions

What you want to know about vector RAG.

Does RAG prevent hallucinations?

It brings references to the answer, but does not guarantee it is correct. Which passages are selected, the instructions and the quality of the sources remain decisive. Check important facts in the documents cited.

Does the wiki replace my documents?

No. The notes are a reading and summary layer. The originals are what let you check a wording, a date or an interpretation.

Does the wiki learn automatically from every conversation?

A conversation and a validated piece of knowledge are two different things. To keep an analysis in the wiki, there has to be an explicit mechanism for folding it back in and checking it; not every answer should become a reference.

How do we handle contradictory sources?

Keep their dates and their references. The system can help spot the discrepancies, but someone has to decide which version governs. Flagging a contradiction is not the same as resolving it.

To go further

The ideas behind the mechanisms.

RAG — Lewis et al., 2020LLM Wiki — a proposal by Andrej Karpathy

The LLM Wiki is a principle for organising knowledge, not a standard imposing one implementation.

Let us see what your sources
can tell you.

A real corpus and a few questions are enough to frame a demonstration.

Book a meeting

RDV démo

Ouvrir l’agenda dans un nouvel onglet ↗
Besoin d'aide ?