The LLM is the bottleneck. Building the graph costs two LLM calls per
chunk (graph extraction and summarization), and almost every other tuning
question comes down to how many of those calls run, how fast, and on which
model. Database and embedding work is small next to that.
Use it
1. Estimate before you ingest
estimate = await cognee.remember("./docs", dry_run=True)
print(estimate) # per-stage token counts and approximate cost; no LLM calls
The estimate covers graph extraction and summarization only, not
improve(), embeddings, or contradiction detection. Not available with
GLiNER or a remote instance.
2. Ingestion knobs
All are remember() arguments (and cognify() ones):
| Knob | Default | What it controls |
|---|
data_per_batch | 20 | How many documents are processed at the same time (a concurrency limit, not a batch). Lower it to reduce load; raise it for many small files. |
chunks_per_batch | 2000 (env CHUNKS_PER_BATCH) | How many chunks each extraction/storage call receives. Lower it to spread load and fail smaller. The DLT pipeline defaults to 100. |
chunk_size | min(embedding max tokens, LLM max tokens / 2); about 8191 tokens with default OpenAI models | Max tokens per chunk. Fewer, bigger chunks mean fewer LLM calls; smaller chunks mean finer-grained graphs. |
run_in_background | False | Return immediately (status="running"); await result later. |
self_improvement | True | improve() after the graph is built. False (or IMPROVE_AUTO_ENABLED=false) skips that extra work. |
Worst-case concurrent LLM calls are about
data_per_batch × min(chunks per document, chunks_per_batch) × 2, and there
is no separate concurrency cap. The rate limiter below is the brake.
Several datasets in one cognify(datasets=[...]) call run one after
another, not in parallel.
3. LLM throttling and models
| Env var | Default | Notes |
|---|
LLM_RATE_LIMIT_ENABLED | false | Turn on a requests-per-interval limit |
LLM_RATE_LIMIT_REQUESTS / LLM_RATE_LIMIT_INTERVAL | 60 / 60 s | 10 requests for local providers (Ollama, llama.cpp, LM Studio) |
AUTO_RATE_LIMIT | true | On a 429/503/529 or timeout, switches the limiter on for 15 minutes (extended while errors continue) |
LLM_EXTRACTION_MODEL / _PROVIDER / _ENDPOINT / _API_KEY / _API_VERSION | the main model | Use a cheaper or faster model just for graph extraction |
LLM_SUMMARIZATION_*, LLM_QUERY_* | the main model | Same, for summaries and for answering queries |
LLM_MAX_COMPLETION_TOKENS | 16384 | Also caps the default chunk size (half of it) |
LLM_ARGS | — | Extra litellm arguments merged into every call, e.g. a request timeout |
Structured-output calls retry with exponential backoff for at least two
attempts and four minutes before failing; auth, not-found, and quota errors
are not retried.
The rate limiter is built from the main model's settings. A stage routed to
a local model does not get the local default of 10 requests; set
LLM_RATE_LIMIT_REQUESTS yourself.
4. Embeddings
| Env var | Default |
|---|
EMBEDDING_BATCH_SIZE | 36 texts per request |
EMBEDDING_MAX_CONCURRENT_DATA_POINTS | 150 (so 150 / 36 = 4 concurrent requests) |
EMBEDDING_RATE_LIMIT_ENABLED / _REQUESTS / _INTERVAL | off / 60 / 60 s |
Embedding calls retry transient errors, rate-limit 429s included, with
exponential jitter for up to 128 s (budget-exhaustion errors are terminal,
not retried), but AUTO_RATE_LIMIT does not apply to them; enable
EMBEDDING_RATE_LIMIT_ENABLED yourself for sustained load. Without any
credentials cognee embeds locally with fastembed (BAAI/bge-small-en-v1.5),
which runs on the CPU and blocks while it works.
remember(data, extractor="gliner") builds the graph and summaries with a
local GLiNER model: no LLM calls, embeddings still run. The runtime installs
on first use (GLINER_AUTO_INSTALL=false to disable; then pip install "cognee[gliner]"). See the cognee-ingestion skill for its limits.
6. Read latency
AUTO_FEEDBACK=false is the biggest win for chat-style use: by default
every answered turn (the default session is used when no session_id is
passed) makes one extra LLM call to analyze feedback. Keep CACHING=true so session memory still works.- Pick a type without an LLM when you need passages, not an answer:
CHUNKS, CHUNKS_LEXICAL, SUMMARIES, CODE, SKILLS. Completion types
make one LLM call; GRAPH_COMPLETION_COT, _DECOMPOSITION,
_CONTEXT_EXTENSION, GRAPH_SUMMARY_COMPLETION and TEMPORAL make
several; NATURAL_LANGUAGE makes one and retries (up to 3 attempts) only
on an empty or failed query; FEELING_LUCKY adds one to pick the type.
- Narrow the search: pass
datasets=[...] (one search runs per dataset)
and a smaller top_k (per dataset, default 15; HYBRID caps each lane at
10, so only values below 10 shrink its context).
SESSION_SEARCH_MODE=concurrent (default) overlaps the feedback analysis
with the answer; sequential runs them back to back.
7. Several datasets at once
With access control on (the default), each dataset has its own databases.
DATASET_QUEUE_MAX_CONCURRENT (default 6, from DATABASE_MAX_LRU_CACHE_SIZE)
caps how many datasets are processed at once in one process, and
SUBPROCESS_IDLE_TTL_SECONDS (600) keeps idle database workers warm.
DATASET_QUEUE_ENABLED (default true) enforces that cap, releases
subprocess engines when a dataset's last scope exits (they stay warm for
SUBPROCESS_IDLE_TTL_SECONDS and close at once only when it is 0) and pins
in-use engines against eviction.
Setting it false removes the cap rather than disabling parallelism, and
risks file-lock leaks and engine eviction under parallel load, so keep it
on.
8. Scaling a deployment
Open-source cognee scales vertically, in one process:
- The API server runs one worker (
gunicorn -w 1), and the improve and
session locks, caches and semaphores are in-process only. More workers or
replicas against the same data are not coordinated (the Helm chart README
says to validate before scaling past one replica).
- The default embedded databases (SQLite, LanceDB, Ladybug) live on local
disk. For production, use Postgres + PGVector and a graph-native backend
such as Neo4j.
distributed/deploy/ holds one-click deploy templates (Modal, Fly,
Railway, Render, Daytona), not a distributed runner.
For production scale: horizontal scaling (distributed ingestion across
workers), cognee's enterprise GLiNER extraction, and the production
Postgres graph adapter are part of cognee's proprietary offering. Contact
social@cognee.ai.
Pitfalls
CHUNK_SIZE, CHUNK_OVERLAP, CHUNK_STRATEGY (in .env.template) and
cognee.config.set_chunk_size() etc. have no effect on ingestion. Pass
chunk_size= to remember() instead.LLM_RATE_LIMIT_TOKENS and EMBEDDING_RATE_LIMIT_TOKENS do nothing;
only request-count limits are enforced.FALLBACK_MODEL is used only when the main model rejects content on
policy grounds, not when it is slow, down, or rate-limited.data_per_batch=None crashes. Omit the argument to get the default.- A background
cognify() is not tracked by
cognee.wait_for_background_tasks() (a background remember() or
improve() is). Await its pipeline status yourself.
- Big default chunks with a small local embedder. The default chunk size
follows the configured embedding limit (8191), even for fastembed's
512-token model. Pass a smaller
chunk_size for local embedding models.
- Benchmark with
CACHING=true: with it off you measure cognee without its
memory layer.
How it works
remember() → add() → cognify(): every document of the dataset is
scheduled at once, data_per_batch of them run concurrently, and each runs
the task chain (classify, chunk, extract graph + summarize, store). Tasks
receive their input in batches of chunks_per_batch.
extract_graph_and_summarize runs extraction and summarization together,
each gathering one LLM call per chunk in the batch. Every LLM call passes
through one process-wide rate limiter.
- Cognify task list and defaults:
cognee/api/v1/cognify/cognify.py
- Concurrency (
data_per_batch semaphore): cognee/modules/pipelines/operations/run_tasks.py
- Task batching:
cognee/modules/pipelines/tasks/task.py, run_tasks_base.py
- LLM calls per chunk:
cognee/tasks/graph/extract_graph_and_summarize.py
- Chunk-size default:
cognee/infrastructure/llm/utils.py:get_max_chunk_tokens
- LLM config, rate limits, stage routing:
cognee/infrastructure/llm/config.py,
cognee/shared/rate_limiting.py, cognee/infrastructure/llm/overload_policy.py,
cognee/infrastructure/llm/retry_config.py
- Embedding config:
cognee/infrastructure/databases/vector/embeddings/config.py
- Dataset queue:
cognee/infrastructure/databases/dataset_queue/queue.py
- Cost estimate:
cognee/modules/cognify/estimator.py
Extending it
- Benchmarks:
cognee/tests/performance/ (batch_add_cognify_test.py,
locust_performance_analysis.py, and statistics_percentile/ for p50–p99
runs with mocked or replayed LLM calls). Measure before and after a change.
- Retrieval quality and cost:
cognee/eval_framework/ (python -m cognee.eval_framework; Modal runs cost money).
- A new pipeline task that calls an LLM should keep
needs_llm=True and go
through LLMGateway, so it shares the rate limiter and retry policy.
anthropics/skills180k
