Feature flags - cognis

Cognis is feature-light by default. Every external integration — providers, vector stores, checkpoint backends, observability exporters — is feature-gated so the core crates compile with no network deps. Pick what you need; pay for nothing else.

cognis (umbrella)

Feature	Pulls in
`openai` (default)	OpenAI provider + embeddings
`ollama` (default)	Ollama provider + embeddings
`anthropic`	Anthropic Messages
`google`	Gemini chat + embeddings
`azure`	Azure OpenAI
`voyage`	Voyage embeddings
`all-providers`	Every provider above
`cache-sqlite`	SQLite-backed model cache
`tools-http`	HTTP request primitives for tools

cargo add cognis --features anthropic,google,vectorstore-faiss
cargo add cognis --features all-providers

cognis-core, cognis-macros

These crates compile with no network features ever. Anything that does I/O lives in a sibling crate.

cognis-llm

Feature	Effect
`openai` (default)	`reqwest` + OpenAI client
`ollama` (default)	`reqwest` + Ollama client (no key)
`anthropic`	`reqwest` + Anthropic client
`google`	`reqwest` + Gemini client
`azure`	`reqwest` + Azure OpenAI client
`all-providers`	All of the above

OpenRouter uses the OpenAI client with Provider::OpenRouter; no separate feature.

cognis-rag

Embeddings

Feature	Provider
`openai` (default)	OpenAI text-embedding-3-*
`ollama` (default)	Local Ollama embeddings
`google`	Google text-embedding-004, gemini-embedding
`voyage`	Voyage AI

Vector stores

Feature	Backend
`vectorstore-faiss`	Local FAISS index (no HTTP)
`vectorstore-chroma`	Chroma server
`vectorstore-qdrant`	Qdrant server
`vectorstore-pinecone`	Pinecone (managed)
`vectorstore-weaviate`	Weaviate server
`all-vectorstores`	All of the above

Loaders

Feature	Loader
`csv-loader`	CSV
`html-loader`	HTML
`yaml-loader`	YAML
`toml-loader`	TOML
`web-loader`	HTTP fetch
`pdf-loader`	PDF
`all-loaders`	All loaders

InMemoryVectorStore, FakeEmbeddings, CachedEmbeddings, BatchedEmbeddings, the indexing pipeline, and the in-memory record manager are always available.

cognis-graph

Feature	Pulls in
`sqlite`	`sqlx/sqlite` for `SqliteCheckpointer`
`postgres`	`sqlx/postgres` for `PostgresCheckpointer`
`serializer-cbor`	CBOR-encoded checkpoint payloads

InMemoryCheckpointer and the engine are always available.

cognis-trace

Feature	Pulls in
`stdout` (default)	`StdoutExporter`
`langfuse`	`reqwest`, `secrecy`, `base64`, plus Langfuse exporter / prompts / scorer
`all`	Every exporter
`integration_tests`	Tests against real services (off by default)

MockExporter is always built — useful in tests.

Picking the minimum set

Local Ollama only
OpenAI + RAG with FAISS
Multi-provider with traces
Stateful agents on Postgres

cognis = { version = "0.3", features = ["ollama"] }

cognis = { version = "0.3", features = ["openai"] }
cognis-rag = { version = "0.3", features = ["vectorstore-faiss", "openai"] }

cognis = { version = "0.3", features = ["all-providers", "cache-sqlite"] }
cognis-trace = { version = "0.3", features = ["langfuse"] }

cognis = { version = "0.3", features = ["openai"] }
cognis-graph = { version = "0.3", features = ["postgres"] }
cognis-trace = { version = "0.3", features = ["langfuse"] }

Installation

Where features live in your Cargo.toml.

Environment variables

What Client::from_env() reads.

​cognis (umbrella)

​cognis-core, cognis-macros

​cognis-llm

​cognis-rag

​Embeddings

​Vector stores

​Loaders

​cognis-graph

​cognis-trace

​Picking the minimum set

​See also