search HTTP API return "the knowledge-graph context is empty"

[LStromann]

=1. Added data with cognee-cli data data_dir && cognee-cli cognify
2. run the backend server with ENABLE_BACKEND_ACCESS_CONTROL=false REQUIRE_AUTHENTICATION=false cognee-cli -ui
3. query with curl command

curl --request POST \
  --url http://localhost:8000/api/v1/search \
  --header 'Content-Type: application/json' \
  --data '
{
  "searchType": "GRAPH_COMPLETION",
  "datasets": [
    "main_dataset"
  ],
  "query": "What is in the document?",
  "topK": 1
}
'

and it returns ["The document contains no nodes or connections — the provided knowledge-graph context is empty."]

However with cognee-cli search "What is in the document?" returns with

% cognee-cli search "What is in the document?"
...
Response: {'search_result': ['...some meaningful txt']}

What do I have to do to get the proper search result with HTTP API?

(using cognee==0.5.0)

This discussion was automatically pulled from Discord.

[LStromann]

When you added data with cognee-cli was ENABLE_BACKEND_ACCESS_CONTROL set to false?

[LStromann]

or only for the backend server

[LStromann]

Because adding and cognifying data also must be run with ENABLE_BACKEND_ACCESS_CONTROL set to false (we use separate storage systems depending if the mode is turned on or off)

[LStromann]

and with Cognee v0.5.0 ENABLE_BACKEND_ACCESS_CONTROL=true is the default

[LStromann]

Yep, that must be the reason. Thanks.

And,, do you know what the recommended approach is when a lock issue occurs?
For example, is it acceptable to delete the lock file shown in the logs as a temporary workaround?
2025-12-16T12:44:37.032151 [error ] Failed to initialize Kuzu database: IO exception: Could not set lock on file : /Users/jeesoo/workspace/rag/.venv/lib/python3.12/site-packages/cognee/.cognee_system/databases/cognee_graph_kuzu

[LStromann]

adding ENABLE_BACKEND_ACCESS_CONTROL=false in add script is not seems to work.

1. first I clean the data

ENABLE_BACKEND_ACCESS_CONTROL=false \
REQUIRE_AUTHENTICATION=false \
cognee-cli delete --all --force

2. Add data

ENABLE_BACKEND_ACCESS_CONTROL=false \
REQUIRE_AUTHENTICATION=false \
cognee-cli add sample_data

ENABLE_BACKEND_ACCESS_CONTROL=false \
REQUIRE_AUTHENTICATION=false \
cognee-cli cognify

3. restart the server

ENABLE_BACKEND_ACCESS_CONTROL=false \
REQUIRE_AUTHENTICATION=false \
cognee-cli -ui

4. Query

curl --request POST \
  --url http://localhost:8000/api/v1/search \
  --header 'Content-Type: application/json' \
  --data '
{
  "searchType": "GRAPH_COMPLETION",
  "datasets": [
    "main_dataset"
  ],
  "query": "What is in the document?",
  "topK": 1
}
'

and return is ["No information available in the provided context (no nodes or relationships), so I cannot determine what is in the document."]

[LStromann]

Best way to resolve the issue is to just run the Cognee backend server and all requests through it (add cognify and etc.)

As there would be just one Kuzu adapter that communicates with the database and handle the lock files

If you need to run it with the cognee cli you can use the redis lock mechanism: https://docs.cognee.ai/core-concepts/sessions-and-caching

So all requests for Kuzu would go through a redis queue but that is a bit more complicated to setup

[LStromann]

I will try it out and check

[LStromann]

Thanks! As suggested, I’ll refactor my scripts to use a unified HTTP API and will update you if the issue is resolved.

[LStromann]

<@508682252595626015> I assume we can't use kuzu lock from pure sdk?

[LStromann]

In this case we can't as the cognee-cli spawns multiple processes of Cognee that have different Kuzu adapter handlers and it has to be a singleton to avoid the locks, so we can either handle it with Laszlo's Redis Queue as it handles read/write operations for kuzu as a singleton (no two requests can come at the same time to the Kuzu DB) or by having users use Cognee as a backend to handle all requests as then Kuzu is handled through our singleton adapter

[LStromann]

But just using one Cognee instance with SDK that is also fine, the problem is that cognee-cli spawns new Cognee instances

If we can get cognee-cli to reuse cognee processes that would work