Vector stores¶

RAGMill talks to every backend through one interface, BaseVectorStore, so your code doesn't change when you switch stores — only environment variables do. Pick the store with RAGMILL_STORE_TYPE and store_from_config() returns the right implementation.

Backend	`RAGMILL_STORE_TYPE`	Best for
SQLite (default)	`sqlite`	local dev, single-folder knowledge bases, offline use
Qdrant	`qdrant`	self-hosted or managed cloud, larger corpora, a built-in vector dashboard
Pinecone	`pinecone`	fully managed cloud, scale

from ragmill.config import RAGMillConfig
from ragmill.vector_store import store_from_config

store = store_from_config(RAGMillConfig.from_env())   # type decided by env

SQLite (local, default)¶

Zero setup. It's a single file (or in-memory) with brute-force dot-product search — fast for thousands of chunks.

export RAGMILL_STORE_TYPE=sqlite
export RAGMILL_SQLITE_PATH=./ragmill.db     # omit → in-memory, not persisted

from ragmill.vector_store import SQLiteVectorStore
store = SQLiteVectorStore("ragmill.db")

Scaling limit

The SQLite search is O(n) per query — great to ~tens of thousands of chunks, not for millions. Past that, use Qdrant/Pinecone (approximate nearest neighbor) behind the same search() interface.

Qdrant¶

pip install "ragmill[qdrant]"
export RAGMILL_STORE_TYPE=qdrant
export RAGMILL_QDRANT_URL=http://localhost:6333     # or your Qdrant Cloud URL
export RAGMILL_QDRANT_API_KEY=your-key              # required for Qdrant Cloud
export RAGMILL_QDRANT_COLLECTION_NAME=ragmill

Run Qdrant locally in one command:

docker run -p 6333:6333 qdrant/qdrant

Payload indexes for filename and source_file are created automatically the first time the collection is set up (Qdrant Cloud requires them for filtered search/delete/sync).

Built-in visualization

Qdrant ships a web dashboard at http://localhost:6333/dashboard (or your cloud cluster's dashboard) that lists collections and visualizes your vectors — the easiest way to see your embedded data.

Pinecone¶

pip install "ragmill[pinecone]"
export RAGMILL_STORE_TYPE=pinecone
export RAGMILL_PINECONE_API_KEY=your-key
export RAGMILL_PINECONE_ENVIRONMENT=us-west1-gcp
export RAGMILL_PINECONE_INDEX_NAME=ragmill

Manage and inspect indexes from the Pinecone console.

Visualizing a local SQLite store¶

The SQLite store keeps text and metadata in a chunks table; embeddings are raw float32 BLOBs.

Browse text/metadata: open ragmill.db in DB Browser for SQLite. The embedding column is opaque there.
Visualize the vectors: either migrate to Qdrant (below) and use its dashboard, or project the 384-dim vectors to 2-D with UMAP/t-SNE in a notebook and plot with Plotly. store.scroll() pages through every record including its embedding for exactly this.

To move a local store into a cloud one, see Migrating backends.