Integrations

Synaptic provides optional integration crates that connect to external services. Each integration is gated behind a Cargo feature flag and adds no overhead when not enabled.

Available Integrations

Integration	Feature	Purpose
OpenAI-Compatible Providers	`openai`	Groq, DeepSeek, Fireworks, Together, xAI, MistralAI, HuggingFace, Cohere, OpenRouter
Azure OpenAI	`openai`	Azure-hosted OpenAI models (chat + embeddings)
Anthropic	`anthropic`	Anthropic Claude models (chat + streaming + tool calling)
Google Gemini	`gemini`	Google Gemini models via Generative Language API
Ollama	`ollama`	Local LLM inference with Ollama (chat + embeddings)
AWS Bedrock	`bedrock`	AWS Bedrock foundation models (Claude, Llama, Mistral, etc.)
Cohere Reranker	`cohere`	Document reranking for improved retrieval quality
Qdrant	`qdrant`	Vector store backed by the Qdrant vector database
PgVector	`pgvector`	Vector store backed by PostgreSQL with the pgvector extension
Pinecone	`pinecone`	Managed vector store backed by Pinecone
Chroma	`chroma`	Open-source vector store backed by Chroma
MongoDB Atlas	`mongodb`	Vector search backed by MongoDB Atlas
Elasticsearch	`elasticsearch`	Vector store backed by Elasticsearch kNN
Redis	`redis`	Key-value store and LLM response cache backed by Redis
SQLite Cache	`sqlite`	Persistent LLM response cache backed by SQLite
PDF Loader	`pdf`	Document loader for PDF files
Tavily Search	`tavily`	Web search tool for agents

Enabling integrations

Add the desired feature flags to your Cargo.toml:

[dependencies]
synaptic = { version = "0.3", features = ["openai", "qdrant", "redis"] }

You can combine any number of feature flags. Each integration pulls in only the dependencies it needs.

Trait compatibility

Every integration implements a core Synaptic trait, so it plugs directly into the existing framework:

OpenAI-Compatible, Azure OpenAI, and Bedrock implement ChatModel -- use them anywhere a model is accepted.
OpenAI-Compatible (MistralAI, HuggingFace, Cohere) and Azure OpenAI also implement Embeddings.
Cohere Reranker implements DocumentCompressor -- use it with ContextualCompressionRetriever for two-stage retrieval.
Qdrant, PgVector, Pinecone, Chroma, MongoDB Atlas, and Elasticsearch implement VectorStore -- use them with VectorStoreRetriever or any component that accepts &dyn VectorStore.
Redis Store implements Store -- use it anywhere InMemoryStore is used, including agent ToolRuntime injection.
Redis Cache and SQLite Cache implement LlmCache -- wrap any ChatModel with CachedChatModel for persistent response caching.
PDF Loader implements Loader -- use it in RAG pipelines alongside TextSplitter, Embeddings, and VectorStore.
Tavily Search implements Tool -- register it with an agent for web search capabilities.

Guides

LLM Providers

OpenAI-Compatible Providers -- Groq, DeepSeek, Fireworks, Together, xAI, MistralAI, HuggingFace, Cohere, OpenRouter
Azure OpenAI -- Azure-hosted OpenAI models
Anthropic -- Anthropic Claude models
Google Gemini -- Google Gemini models
Ollama -- Local LLM inference (chat + embeddings)
AWS Bedrock -- AWS Bedrock foundation models

Reranking

Cohere Reranker -- document reranking for improved retrieval

Vector Stores

Qdrant Vector Store -- store and search embeddings with Qdrant
PgVector -- store and search embeddings with PostgreSQL + pgvector
Pinecone Vector Store -- managed vector store with Pinecone
Chroma Vector Store -- open-source embedding database
MongoDB Atlas Vector Search -- vector search with MongoDB Atlas
Elasticsearch Vector Store -- vector search with Elasticsearch kNN

Storage & Caching

Redis Store & Cache -- persistent key-value storage and LLM caching with Redis
SQLite Cache -- local LLM response caching with SQLite

Loaders & Tools

PDF Loader -- load documents from PDF files
Tavily Search Tool -- web search tool for agents

Keyboard shortcuts

Synaptic