This page is machine-translated from French. Read the French original.

Vector base and embeddings

These two elements determine the quality of the documentary research. They are also the two most expensive choices to change afterward.

The mechanism

1. Each document fragment is transformed into vector by the embeddings engine — a list of numbers representing its meaning.
2. These vectors are stored in the vector base.
3. The question is vectorized by the same engine, then compared to the stored vectors.
4. The closest fragments are selected and provided to the model.

Hence the central constraint: All vectors in an index must originate from the same embeddings engine.. Two different engines produce incomparable performances.

The embeddings engine

Réglages → Préférence d'embeddings.

integrated motor (by default) — runs locally, requires no key, no network output. Suitable for the vast majority of uses, including multilingual.

External motors — Several providers are supported. They can offer advantages for demanding or highly specialized multilingual corpora, at the cost of a network output. with each indexing and each question, and a cost per use.

Changing the engine invalidates all existing indexes. No error message will indicate this: search results will simply become inaccurate. All search areas must be reindexed. Decide this at startup.

Text segmentation

Réglages → Découpage du texte two parameters are set:

• Fragment size — shorter, more precise but less contextual; longer, the opposite.
• Recovery — the text repeated from one fragment to the next, so as not to split an idea in two.

The default values are suitable for most corpora. A change only affects documents indexed later.

When to adjust: Highly structured and short documents (fact sheets, questions and answers) benefit from smaller fragments; long argumentative texts (contracts, case law) benefit from larger fragments.

The vector basis

Réglages → Base vectorielle.

Integrated local base (by default) — stored in the installation directory. No service to manage, no network output, backed up with everything else. It easily handles corpora of tens of thousands of fragments, which covers most organizations.

External databases — Several market solutions are supported, either self-hosted or as a managed service. → Full list

When to consider an external base

• The volume exceeds what the local base can comfortably handle.
• A database already exists within your organization and you wish to pool it.
• Several applications must share the same index.
• High availability is required.

In other cases, the local base is the best choice: fewer moving parts, no network output, unified backup.

A managed service means that your vectors and their metadata are outside your scope. For an installation where sovereignty is a requirement, this negates the benefit of a local model.

Index organization

Each workspace has its own own namespace in the database. This ensures that a search in one space cannot reach documents in another.

Each user's Second Brain also has its own separate workspace.

Maintenance operations

Reindex a space. Necessary after a change of embedding engine, or when an index is suspected of being inconsistent. The operation removes the vectors and recalculates them from the documents.

Reset a space. Clears the vectors from the space. The documents remain in the library and must be added again.

Check the number of vectors. The administration displays the indexed total — useful for confirming that an indexation has been successful.

Diagnosing a bad search

The thresholds and the number of fragments are adjustable. per workspace. → Conversations