Databases (Vector Stores) Reference

Vector databases store embeddings and enable semantic search. LlamaFarm supports multiple vector store backends for different deployment scenarios.

Quick Start

Databases are configured in rag.databases:

rag:
  default_database: main_db
  databases:
    - name: main_db
      type: ChromaStore
      config:
        collection_name: documents
        distance_function: cosine

Common Database Properties

Property	Required	Description
name	Yes	Unique identifier (lowercase, underscores)
type	Yes	Store type (ChromaStore, QdrantStore, etc.)
config	No	Store-specific configuration
embedding_strategies	No	List of embedding configurations
retrieval_strategies	No	List of retrieval configurations
default_embedding_strategy	No	Default embedder name
default_retrieval_strategy	No	Default retrieval strategy name

---

ChromaStore

Chroma is a lightweight, embedded vector database. Great for development and small-to-medium deployments.

Best for: Local development, prototyping, embedded use

- name: main_db
  type: ChromaStore
  config:
    collection_name: documents
    distance_function: cosine
    persist_directory: ./data/chroma_db

Options

Option	Type	Default	Description
collection_name	string	documents	Collection name (alphanumeric, hyphens, underscores)
host	string	null	Server host (null for embedded)
port	integer	8000	Server port
distance_function	string	cosine	cosine, l2, ip
distance_metric	string	cosine	Alternative name for distance_function
embedding_dimension	integer	768	Vector dimension (1-4096)
enable_deduplication	boolean	true	Enable document deduplication
embedding_function	string	null	Built-in embedding function

Distance Functions

Function	Description	Use Case
cosine	Cosine similarity	General purpose, normalized vectors
l2	Euclidean distance	When magnitude matters
ip	Inner product	Already normalized embeddings

Client vs Server Mode

Embedded (default):

config:
  collection_name: documents
  persist_directory: ./data/chroma

Server mode:

config:
  collection_name: documents
  host: localhost
  port: 8000

---

QdrantStore (Coming Soon)

Coming Soon

QdrantStore is planned but not yet implemented. Use ChromaStore for now.

Qdrant is a high-performance vector database with rich filtering and clustering.

Best for: Production deployments, large datasets, advanced filtering

- name: production_db
  type: QdrantStore
  config:
    collection_name: documents
    host: localhost
    port: 6333
    vector_size: 768
    distance: Cosine

Planned Options

Option	Type	Default	Required	Description
collection_name	string	documents	No	Collection name
host	string	localhost	No	Server host
port	integer	6333	No	REST API port
grpc_port	integer	6334	No	gRPC port
api_key	string	null	No	API key for auth
vector_size	integer	-	Yes	Vector dimension (1-65536)
distance	string	Cosine	No	Cosine, Euclid, Dot
on_disk	boolean	false	No	Store vectors on disk

Distance Metrics

Metric	Description	Use Case
Cosine	Cosine similarity	General semantic search
Euclid	Euclidean distance	Geometric comparisons
Dot	Dot product	Pre-normalized vectors

Docker Setup (for when available)

docker run -p 6333:6333 -p 6334:6334 \
  -v $(pwd)/qdrant_data:/qdrant/storage \
  qdrant/qdrant

---

FAISSStore (Coming Soon)

Coming Soon

FAISSStore is planned but not yet implemented. Use ChromaStore for now.

FAISS (Facebook AI Similarity Search) for high-performance similarity search.

Best for: Large-scale search, research, GPU acceleration

- name: faiss_db
  type: FAISSStore
  config:
    dimension: 768
    index_type: HNSW
    metric: Cosine

Planned Options

Option	Type	Default	Required	Description
dimension	integer	-	Yes	Vector dimension (1-4096)
index_type	string	Flat	No	Flat, IVF, HNSW, LSH
metric	string	L2	No	L2, IP, Cosine
nlist	integer	100	No	Number of clusters (IVF)
nprobe	integer	10	No	Clusters to search (IVF)
use_gpu	boolean	false	No	Enable GPU acceleration

Index Types

Type	Speed	Accuracy	Memory	Best For
Flat	Slow	Exact	High	Small datasets, exact search
IVF	Fast	Good	Medium	Large datasets
HNSW	Very Fast	Very Good	High	Production, low latency
LSH	Fast	Moderate	Low	Very large datasets

---

PineconeStore (Coming Soon)

Coming Soon

PineconeStore is planned but not yet implemented. Use ChromaStore for now.

Pinecone is a fully managed cloud vector database.

Best for: Managed service, enterprise, global scale

- name: cloud_db
  type: PineconeStore
  config:
    api_key: ${PINECONE_API_KEY}
    index_name: my-index
    dimension: 768
    environment: us-east-1-aws

Planned Options

Option	Type	Default	Required	Description
api_key	string	-	Yes	Pinecone API key
index_name	string	-	Yes	Index name (lowercase, hyphens)
dimension	integer	-	Yes	Vector dimension (1-20000)
environment	string	us-east-1-aws	No	Pinecone environment
metric	string	cosine	No	euclidean, cosine, dotproduct
namespace	string	""	No	Namespace for isolation
replicas	integer	1	No	Number of replicas (1-20)

---

Complete Database Configuration

Full example with embedding and retrieval strategies:

rag:
  default_database: main_db

  databases:
    - name: main_db
      type: ChromaStore
      config:
        collection_name: documents
        distance_function: cosine
        persist_directory: ./data/main_db

      # Embedding strategies
      embedding_strategies:
        - name: default
          type: OllamaEmbedder
          config:
            model: nomic-embed-text
            dimension: 768
          priority: 0

        - name: openai
          type: OpenAIEmbedder
          config:
            model: text-embedding-3-small
            api_key: ${OPENAI_API_KEY}
          priority: 1

      # Retrieval strategies
      retrieval_strategies:
        - name: semantic
          type: BasicSimilarityStrategy
          config:
            distance_metric: cosine
            top_k: 10
          default: true

        - name: filtered
          type: MetadataFilteredStrategy
          config:
            top_k: 10
            filter_mode: pre

      default_embedding_strategy: default
      default_retrieval_strategy: semantic

Multiple Databases

Configure separate databases for different use cases:

databases:
  # Primary document store
  - name: documents_db
    type: ChromaStore
    config:
      collection_name: documents

  # Archive for older content (embedded mode)
  - name: archive_db
    type: ChromaStore
    config:
      collection_name: archive
      persist_directory: ./data/archive

  # Server-mode for production
  - name: production_db
    type: ChromaStore
    config:
      collection_name: production
      host: localhost
      port: 8000

Future Options

When QdrantStore and FAISSStore are available, you'll be able to use them for larger scale deployments and more advanced features.

Query specific databases:

lf rag query --database documents_db "recent reports"
lf rag query --database archive_db "historical data"

Database Selection Guide

Database	Deployment	Scale	Features	Status
ChromaStore	Embedded/Server	Small-Medium	Simple, portable	Available
QdrantStore	Self-hosted	Medium-Large	Rich filtering, clustering	Coming Soon
FAISSStore	Self-hosted	Very Large	GPU support, research	Coming Soon
PineconeStore	Cloud	Any	Managed, global	Coming Soon

Next Steps

• Retrieval Strategies - Configure retrieval methods
• Embedders Reference - Configure embeddings
• RAG Guide - Full RAG overview