fabricatio-milvus 0.3.2

An extension of fabricatio

fabricatio-milvus

MIT

Milvus vector database integration for Fabricatio — store, search, and retrieve document embeddings with the Fabricatio RAG framework.

Installation

pip install fabricatio[milvus]

Or install all Fabricatio components:

pip install fabricatio[full]

Overview

fabricatio-milvus extends Fabricatio's RAG (Retrieval-Augmented Generation) system with a Milvus vector database backend. It provides:

• Pydantic-based document models that auto-generate Milvus collection schemas from field type annotations.
• A MilvusRAG capability class implementing the add_document/afetch_document/aretrieve contract backed by pymilvus.
• Ready-to-use Action subclasses (InjectToDB, MilvusRAGTalk) for building agent pipelines.

Configuration

Milvus settings are loaded from the Fabricatio config system under the milvus namespace:

from fabricatio_milvus.config import milvus_config

# milvus_config.milvus_uri       → str | None
# milvus_config.milvus_timeout   → float | None
# milvus_config.milvus_token     → SecretStr | None
# milvus_config.milvus_dimensions → int | None

These can also be set per-instance via MilvusScopedConfig (pydantic model with the same fields).

Key Components

Models — fabricatio_milvus.models

Class	Description
MilvusDataBase[ST]	Abstract base combining StoredDocumentModel and SearchedDocumentModel. Generates Milvus CollectionSchema from Pydantic fields (int → INT64, str → VARCHAR, float → DOUBLE, list[str] → ARRAY[VARCHAR], JsonValue → JSON). Default index type FLAT, metric type COSINE.
MilvusClassicModel[SD]	Minimal concrete model with a single text: str field.
MilvusScopedConfig	Per-instance override for Milvus connection parameters (uri, token, timeout, dimensions).

Capabilities — fabricatio_milvus.capabilities

Class / Function	Description
create_client(uri, token, timeout)	Cached factory returning a pymilvus.MilvusClient.
AddConfig	Configuration for add_document: collection_name, flush.
FetchConfig[D]	Configuration for afetch_document: document_model, collection_name, similarity_threshold (default 0.37), result_per_query (default 10), tei_endpoint, reranker_threshold, filter_expr.
MilvusRAG[D, AC, FC]	Core RAG class backed by Milvus. Inherits from RAG, MilvusScopedConfig, UseEmbedding, UseReranker, UseLLM.

MilvusRAG methods:

• init_client(milvus_uri, milvus_token, milvus_timeout) — initialize the underlying MilvusClient.
• check_client(init=True) — ensure client is initialized, optionally initializing from config.
• add_document(data, config) — vectorize documents and insert into a collection. Creates the collection if it doesn't exist.
• afetch_document(query, config) — vectorize query strings, search Milvus, deduplicate by ID, sort by distance descending, and deserialize into typed document models.
• aretrieve(query, document_model, ...) — convenience wrapper that builds a FetchConfig and calls afetch_document.

Actions — fabricatio_milvus.actions

Action	Description
InjectToDB	Action that injects MilvusDataBase instances into a Milvus collection. Automatically creates the collection with the correct schema and index if needed. Supports override_inject to drop and recreate.
MilvusRAGTalk	Interactive RAG conversation loop. Queries Milvus with user input, retrieves relevant documents, augments the LLM prompt, and returns generated responses. Runs until the user exits.

Usage Example

from fabricatio_milvus.actions.rag import InjectToDB
from fabricatio_milvus.models.milvus import MilvusClassicModel

# Create document models
docs = [
    MilvusClassicModel(text="Fabricatio is a Python library for building LLM applications."),
    MilvusClassicModel(text="Milvus is an open-source vector database."),
]

# Inject into Milvus
action = InjectToDB(
    collection_name="my_knowledge_base",
    milvus_uri="http://localhost:19530",
)
await action.execute(to_inject=docs)

Dependencies

• fabricatio-core — core interfaces, config, and utilities
• fabricatio-rag — RAG abstractions (RAG, document models)
• pymilvus (≥2.5.4) — Milvus Python SDK
• pydantic (≥2.7.1) — data validation and schema generation
• more-itertools (≥10.8.0) — additional itertools

License

MIT