Configuration

Configuration

Connector process

The connector has no third-party Python dependencies.

Variable / flag	Default	Notes
LLM_BASE_URL	http://127.0.0.1:8000	OpenAI-compatible base, without /v1
LLM_DOCKER_ENDPOINT	http://host.docker.internal:8000/v1	Same server from the Cognee container
LLM_API_KEY	empty	Bearer token for the local server
DISTILL_MODEL / --model	Qwen3.6-35B-A3B MLX alias	Must match /v1/models exactly
DISABLE_THINKING	true	Sends chat_template_kwargs.enable_thinking=false; disable for endpoints that reject extensions
COGNEE_API_URL	http://localhost:8010	Host-side Cognee API
OPENCODE_DB	~/.local/share/opencode/opencode.db	Source SQLite path
--dataset	opencode_kb	Cognee destination dataset
--concurrency	4	Parallel distillation requests
--add-batch-size	100	Documents per idempotent Add request
--data-per-batch	4	Documents concurrently processed by the single Cognify run
--chunks-per-batch	12	Chunks per graph-extraction task batch
--limit, --offset	unset	Smoke tests or manual sharding

Legacy OMLX_URL and OMLX_API_KEY are accepted as fallbacks.

Cognee Docker service

docker-compose.yml exposes:

Service	URL
Cognee API / Swagger	http://localhost:8010 / /docs
Cognee UI	http://localhost:3010

The browser-facing UI URL must match both CORS_ALLOWED_ORIGINS and UI_APP_URL. A mismatch appears as a generic UI connection error even when /health is healthy.

Model routing

Cognee needs both an LLM and an embedding model. The included configuration routes both to an OpenAI-compatible server on the Docker host:

LLM_PROVIDER: custom
LLM_MODEL: openai/<chat-model-id>
LLM_ENDPOINT: http://host.docker.internal:8000/v1
LLM_API_KEY: ${OMLX_API_KEY}

EMBEDDING_PROVIDER: openai_compatible
EMBEDDING_MODEL: <embedding-model-id>
EMBEDDING_ENDPOINT: http://host.docker.internal:8000/v1
EMBEDDING_DIMENSIONS: "1024"  # must match the model

openai/ on LLM_MODEL tells LiteLLM to use the OpenAI protocol. openai_compatible embeddings send the embedding model ID verbatim.

Structured output

Cognee graph extraction is schema-constrained. Local models are configured with:

STRUCTURED_OUTPUT_FRAMEWORK: instructor
LLM_INSTRUCTOR_MODE: json_schema_mode

Reasoning models additionally receive:

LLM_ARGS: '{"chat_template_kwargs": {"enable_thinking": false}}'

See Thinking preamble for why this is a correctness requirement, not cosmetic tuning.

Concurrency guidance

Start at four distill workers and four Cognify documents in flight. Measure completed sessions or documents per hour rather than the UI's single-stream tokens/second.

• Reduce workers when interactive workloads share the same GPU or latency rises continuously.
• Increase carefully when GPU utilization is low and memory pressure is stable.
• Local inference often becomes memory-bandwidth-bound before compute-bound.

Persistence

• var/distilled/: accepted distillation output, required for connector resume.
• var/transcripts/: normalized, scrubbed model input for audit.
• progress.json: distillation state only; Cognee owns ingestion state.
• cognee_storage: Docker named volume containing the knowledge base.
• var/log/: launchd stdout/stderr; container logs are rotated by Compose.