Trying to get simple example with OpenRouter working

[LStromann]

=Hi there... I'm getting strange issues with a simple OpenRouter example.

It's stuck after Switching to TikToken default tokenizer. Anyone with an idea?

LLM_PROVIDER="custom"
LLM_API_KEY="sk-or-v1-d2n8vn423kebabifn90843n934j983443oj"
LLM_ENDPOINT="https://openrouter.ai/api/v1"
LLM_MODEL="openrouter/openai/gpt-5-mini"
LLM_MAX_TOKENS=16384

EMBEDDING_PROVIDER="custom"
EMBEDDING_API_KEY="sk-or-v1-d2n8vn423kebabifn90843n934j983443oj"
EMBEDDING_ENDPOINT="https://openrouter.ai/api/v1/embeddings"
EMBEDDING_MODEL="openrouter/openai/text-embedding-3-large"
EMBEDDING_API_VERSION=""
EMBEDDING_BATCH_SIZE=36
EMBEDDING_DIMENSIONS=3072
EMBEDDING_MAX_TOKENS=8191

cognee-cli --debug add "Cognee turns documents into AI memory."

2026-01-28T03:13:35.685373 [info     ] Logging initialized            [cognee.shared.logging_utils] cognee_version=0.5.1 database_path=/home/dev/tf-project/.venv/lib/python3.12/site-packages/cognee/.cognee_system/databases graph_database_name= os_info='Linux 6.12.63+deb13-amd64 (#1 SMP PREEMPT_DYNAMIC Debian 6.12.63-1 (2025-12-30))' python_version=3.12.12 relational_config=cognee_db structlog_version=25.5.0 vector_config=lancedb

2026-01-28T03:13:35.685483 [info     ] Database storage: /home/dev/tf-project/.venv/lib/python3.12/site-packages/cognee/.cognee_system/databases [cognee.shared.logging_utils]
Note: Debug mode enabled. Full stack traces will be shown.
Adding 1 item(s) to dataset 'main_dataset'...
Processing data...

2026-01-28T03:13:37.621931 [warning  ] Could not get tokenizer from HuggingFace due to: cannot import name 'is_offline_mode' from 'huggingface_hub' (/home/dev/tf-project/.venv/lib/python3.12/site-packages/huggingface_hub/__init__.py) [LiteLLMEmbeddingEngine]

2026-01-28T03:13:37.622086 [info     ] Switching to TikToken default tokenizer. [LiteLLMEmbeddingEngine]

This discussion was automatically pulled from Discord.

[LStromann]

Hey <@721113411848634379>could you update your embedding endpoint?
it should be https://openrouter.ai/api/v1 instead of /embeddings.
Docs are here if you want to double-check: https://docs.cognee.ai/setup-configuration/embedding-providers#custom-providers

let us know if this works

[LStromann]

The docs actually doesn't say much about embeddings with openrouter. Also it didn't find anything about the tokenizer (only one entry in .denv)

[LStromann]

yeah it outlines the env config for a custom provider could you please update and check if this works or not

[LStromann]

also the tokenizer warning you saw is normal and not causing the hang it just falls back to tiktoken if huggingface can't load

[LStromann]

Is there a way to make --debug show actualy connections? Or can I use a proxy? http_proxy doesn't seem to be used.

[LStromann]

But I'll dig around. Thanks for help 🙂

[LStromann]

yep, --debug logs tokenizer and embedding engine actions
for HTTP proxy support you'd need to rely on the underlying httpx library's environment variable handling or implement explicit proxy configuration in the CogneeClient

[LStromann]

Or you can try enabling verbose logging in requests/python to capture api calls