[Bug]: Explicit CHUNKS recall treats identifier-only query as a conversational statement and skips retrieval

[metam1nd]

Summary

An explicit CHUNKS recall with an identifier-only query can be classified as a conversational statement. Cognee then returns an acknowledgement before vector retrieval runs, even though the caller explicitly selected CHUNKS.

Reproduction

Store these two permanent memories in the same dataset:

COGNEE_IMPLICIT_UPDATE_TEST_20260822: The test value is ALPHA.
COGNEE_IMPLICIT_UPDATE_TEST_20260822: The test value is BETA.

Then call the MCP recall tool with:

query: COGNEE_IMPLICIT_UPDATE_TEST_20260822
datasets: main_dataset
search_type: CHUNKS
top_k: 10

Actual behavior

The result is an acknowledgement rather than retrieved chunks:

[graph] Acknowledged: COGNEE_IMPLICIT_UPDATE_TEST_20260822.

The logs confirm that the explicit override was received:

Router override recorded: routed=HYBRID_COMPLETION, user_chose=CHUNKS

However, the normal ChunksRetriever log line for starting chunk retrieval does not appear. Retrieval is skipped.

Expected behavior

Because the caller explicitly selected CHUNKS, Cognee should execute literal chunk retrieval and return matching stored chunks. The expected result includes both:

COGNEE_IMPLICIT_UPDATE_TEST_20260822: The test value is BETA.
COGNEE_IMPLICIT_UPDATE_TEST_20260822: The test value is ALPHA.

Root cause

ChunksRetriever inherits supports_session_turn_preparation = True from BaseRetriever.

In run_sequential_session_turn, prepare_session_turn_for_retrieval runs before get_retrieved_objects. The identifier-only query is interpreted as a conversational statement. When turn preparation returns should_answer = False, the function exits early with an acknowledgement and never invokes ChunksRetriever.get_retrieved_objects.

The current upstream main version of chunks_retriever.py does not opt out of session-turn preparation.

Verified workaround

Adding this class attribute fixes the exact reproduction:

class ChunksRetriever(BaseRetriever):
    supports_session_turn_preparation = False

After applying it, the same MCP call returns the raw BETA and ALPHA chunks, with BETA ranked first. The server and client configuration are otherwise unchanged.

This opt-out pattern already exists for other non-conversational retrievers such as CodeRetriever and GraphReportRetriever.

Environment

• cognee: 1.5.2
• cognee-mcp: 0.5.5
• Docker image digest: sha256:3b300b5d4ab79bc5d03556c41e0613476aa61f14a54903f90b7a815081685ec9
• Image revision: a8f9760
• MCP transport: Streamable HTTP
• Server mode: direct
• Vector and graph stores: LanceDB and Ladybug

Related discussions

• [Bug]: MCP recall concatenates contradictory empty and relevant graph-completion branches #4625 reports a separate MCP formatting bug involving contradictory graph-completion branches.
• Best practices for recall() in an infra ops agent: search type selection, routing, and tradeoffs #3697 discusses recall search-type selection generally but does not describe this early-return behavior.

Would the maintainers accept the ChunksRetriever opt-out above as an upstream fix? It may also be worth reviewing whether other literal retrieval modes should bypass conversational turn preparation.