vector

Lightweight vector module for direct collection management. Agents create named collections, index documents with automatic chunking, and search semantically or with hybrid keyword + vector. Shares the FastEmbed singleton with context_builder and rag - zero extra memory cost when all three are loaded.

Property	Value
Module id	vector
Version	1.0.0
Type	user
Pip deps	fastembed, qdrant-client

Use vector for raw vector ops; use rag for full RAG pipelines (knowledge bases, hybrid retrieval, RRF fusion, citations, semantic cache, Text2SQL). rag reference →

The 14 actions

Mostly risk_level: low (reads), medium for inserts, high for delete_collection.

Tool	Source	Purpose
vector.create_collection		Create a named collection.
vector.delete_collection		Delete a collection + all its docs.
vector.list_collections		List user collections + counts.
vector.add		Add raw text documents (chunks + embeds).
vector.add_file		Read + chunk + embed + add a file.
vector.search		Semantic search.
vector.hybrid_search		Semantic + keyword fusion.
vector.get		Retrieve docs by id.
vector.delete		Delete docs by id (or filter).
vector.update_metadata		Patch metadata on existing docs.
vector.count		Count docs (with optional filter).
vector.collection_stats		Doc count, dimensions, storage info.
vector.add_directory		Walk a directory tree, chunk + embed each file. Skips unchanged files (content-hash dedup).
vector.search_multi		Fan-out semantic search across many collections, merged + re-ranked.

Chunking strategies

Strategy	Splits on	Default size	Best for
fixed	N characters + overlap	500 chars	Structured data, code.
sentence	Sentence boundaries (. ! ?)	500 chars	Natural prose.
paragraph	Double newlines (\n\n)	1000 chars	Articles, docs.
recursive (default)	\n\n → \n → . → → char	500 chars	Universal.

Each chunk carries {text, index, start_char, end_char, metadata}.

Embedding model

Property	Value
Default model	sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 (minilm-l12).
Dimensions	384.
Languages	~50.
Source	FastEmbed (ONNX, CPU).
Sharing	Singleton with context_builder + rag.

Override via config.embedding_model (any FastEmbed-supported HuggingFace id).

Collection isolation

Naming convention: user collections are stored as user_{app_id}_{name} and are kept separate from the system tools collection used by tool discovery. Two apps that pick the same nominal collection name do not clash.

Configuration

tools:
  modules:
    vector:
      config:
        embedding_model: null            # null = use context_builder's singleton
        default_chunk_size: 500
        default_overlap: 50
        persistence_dir: null            # null = in-memory; otherwise Qdrant on-disk path

Aliases (FR / EN)

vector ships aliases for international agents. A few examples:

Action	Aliases
vector.create_collection	creer_collection, new_collection
vector.add	ajouter, indexer, embed, insert
vector.search	rechercher, chercher, query, find_similar
vector.hybrid_search	recherche_hybride
vector.delete	supprimer_documents, remove

Full list per action via @action(aliases=[...]) in

Cross-references

• App-config block reference (tools.modules.vector): App Configuration → tools.modules
• Full RAG pipeline (knowledge bases, RRF, citations, cache, Text2SQL): rag reference
• index module (system-level workspace index): index reference