Reference

Providers

The LLMs, embedders, vector stores, and graph stores ContextKit works with, and how to configure each.

Each component is configured the same way: a provider name and a config dictionary.

memory = Memory.from_config({
    "llm": {"provider": "...", "config": {...}},
    "embedder": {"provider": "...", "config": {...}},
    "vector_store": {"provider": "...", "config": {...}},
})

LLMs

The LLM extracts facts from messages and decides how to update existing memories. It needs to follow instructions and return JSON reliably; small, fast models work well.

`provider`	Key, from the config's `api_key` or the environment	Install
`openai`	`OPENAI_API_KEY` (or `OPENROUTER_API_KEY` to go through OpenRouter)	included
`openai_structured`	`OPENAI_API_KEY`; uses OpenAI structured outputs	included
`azure_openai`, `azure_openai_structured`	`azure_kwargs` or `LLM_AZURE_OPENAI_API_KEY`, `LLM_AZURE_DEPLOYMENT`, `LLM_AZURE_ENDPOINT`, `LLM_AZURE_API_VERSION`	included
`anthropic`	`ANTHROPIC_API_KEY` (see the note below)	`pip install anthropic`
`gemini`	`GOOGLE_API_KEY`	`[llms]`
`groq`	`GROQ_API_KEY`	`[llms]`
`together`	`TOGETHER_API_KEY`	`[llms]`
`deepseek`	`DEEPSEEK_API_KEY`	included
`xai`	`XAI_API_KEY`	included
`ollama`	none; `ollama_base_url` (default `http://localhost:11434`)	`[llms]`
`lmstudio`	none; `lmstudio_base_url`	included
`vllm`	`VLLM_API_KEY`; `vllm_base_url` or `VLLM_BASE_URL`	included
`litellm`	Whatever the LiteLLM model needs	`[llms]`
`aws_bedrock`	AWS credentials from the environment	`[extras]`
`sarvam`	`SARVAM_API_KEY`	included
`langchain`	Pass a LangChain chat model object as `model`	`pip install langchain`

Common config keys: model, temperature (0.1), max_tokens (2000), top_p (0.1), and api_key. Always set model: some providers' built-in defaults are large or retired models.

memory = Memory.from_config({
    "llm": {"provider": "groq", "config": {"model": "llama-3.3-70b-versatile", "temperature": 0.1}},
})

Anthropic models

Up to 0.1.2, the anthropic provider sends both temperature and top_p with every request, which every Claude model available today rejects. From 0.1.3 it sends only temperature, which Claude Haiku 4.5, Sonnet 4.6, and Opus 4.6 accept, for example "model": "claude-sonnet-4-6". Newer Claude models reject temperature as well, so they do not work with this provider yet. Its built-in default model has been retired, so always set model.

Embedders

The embedder turns memories and queries into vectors. Its output size must match the vector store's embedding_model_dims; see matching the embedding size.

`provider`	Default model	Size	Install
`openai`	`text-embedding-3-small`	1536	included
`azure_openai`	your deployment; keys in `azure_kwargs` or `EMBEDDING_AZURE_*`		included
`ollama`	`nomic-embed-text`	set `embedding_dims` to the model's size (768 for `nomic-embed-text`)	`[llms]`
`huggingface`	`multi-qa-MiniLM-L6-cos-v1` (sentence-transformers)	read from the model	`[extras]`
`gemini`	`models/text-embedding-004`	768	`[llms]`
`vertexai`	`text-embedding-004`		`[llms]`
`together`			`[llms]`
`lmstudio`			included
`aws_bedrock`			`[extras]`
`langchain`	Pass a LangChain embeddings object as `model`		`pip install langchain`

Vector stores

`provider`	Typical `config`	Install
`qdrant` (default)	`path` + `on_disk: True` locally, `host` + `port` for a server, or `url` + `api_key` for Qdrant Cloud	included
`pgvector`	`dbname`, `user`, `password`, `host`, `port`	`pip install psycopg2-binary`
`chroma`	`path`, or `host` + `port`	`[vector_stores]`
`pinecone`	`api_key`, `serverless_config` or `pod_config`	`[vector_stores]`
`weaviate`	`cluster_url`, `auth_client_secret`	`[vector_stores]`
`faiss`	`path`	`[vector_stores]`
`milvus`	`url`, `token`	`pip install pymilvus`
`mongodb`	`mongo_uri`, `db_name`	`[vector_stores]`
`redis`	`redis_url`	`pip install redis redisvl`
`elasticsearch`	`host`, `port`, or `cloud_id` + `api_key`	`[extras]`
`opensearch`	`host`, `port`, `http_auth`	`[extras]`
`supabase`	`connection_string`	`[vector_stores]`
`upstash_vector`	`url`, `token`	`[vector_stores]`
`azure_ai_search`	`service_name`, `api_key`	`[vector_stores]`
`vertex_ai_vector_search`	`project_id`, `endpoint_id`, `index_id`, `deployment_index_id`	`[llms]`
`langchain`	Pass a LangChain vector store as `client`	`pip install langchain`

Every vector store also takes collection_name (default outhad_contextkit) and embedding_model_dims (default 1536). The configuration reference and each provider's config class have the full list.

memory = Memory.from_config({
    "vector_store": {
        "provider": "pgvector",
        "config": {"dbname": "app", "user": "app", "password": "...", "host": "localhost", "port": 5432},
    },
})

Graph stores

`provider`	`config`	Install
`neo4j`	`url`, `username`, `password`, optional `database`	`[graph]`
`memgraph`	`url`, `username`, `password`	`[extras]`
`neptune`	`endpoint="neptune-graph://<graph-id>"`	`[graph]`

See Graph memory.

Run everything locally

Ollama can serve both the LLM and the embedder, so memories never leave your machine:

pip install "outhad-contextkitai[llms]"
ollama pull llama3.1:8b
ollama pull nomic-embed-text

from outhad_contextkit import Memory

memory = Memory.from_config({
    "llm": {"provider": "ollama", "config": {"model": "llama3.1:8b", "temperature": 0}},
    "embedder": {"provider": "ollama", "config": {"model": "nomic-embed-text", "embedding_dims": 768}},
    "vector_store": {
        "provider": "qdrant",
        "config": {"path": "./memory", "on_disk": True, "embedding_model_dims": 768},
    },
})

Edit this page on GitHub