[BUG] Qdrant metadata content overwrites the indexed document body #9129

Description

Qdrant document metadata containing a content key silently replaces the document text in the stored payload. The embedding is computed from document["content"], but search returns metadata["content"] as the document body. Both add_documents and aadd_documents use the same payload helper.

Reproduction

Reproduced on main at 6bd60b223101a11f7ac675706ff1ebb03314c95d (CrewAI 1.15.35), Windows, Python 3.12.15, qdrant-client 1.14.3. No server, model, or API key is needed:

from crewai.rag.qdrant.client import QdrantClient
from qdrant_client import QdrantClient as SDKClient
from qdrant_client.models import Distance, VectorParams

sdk = SDKClient(location=":memory:")
try:
    client = QdrantClient(client=sdk, embedding_function=lambda text: [1.0, 0.0, 0.0])
    client.create_collection(
        collection_name="content_repro",
        vectors_config=VectorParams(size=3, distance=Distance.COSINE),
    )
    client.add_documents(collection_name="content_repro", documents=[{
        "content": "Original document text",
        "metadata": {"content": "Metadata text", "source": "example.txt"},
    }])
    result = client.search(collection_name="content_repro", query="Original document text")[0]
    print(result["content"])
    assert result["content"] == "Original document text", result
finally:
    sdk.close()

Actual: prints Metadata text and fails the assertion. The equivalent async create/add/search sequence returns the same wrong text.

Expected: the canonical document body remains Original document text. Non-conflicting metadata such as source remains available.

Cause and proposed fix

_create_point_from_document builds {"content": doc["content"], **metadata}, so metadata wins on collision. Assigning the reserved payload content field after expanding metadata preserves the document body. This keeps the existing flat payload layout and metadata filtering behavior; it does not introduce a new metadata storage format or recover previously overwritten content.

I checked existing issues and open PRs for this collision and did not find a duplicate. A one-line fix and regression tests for mapping/list metadata and empty/nonempty text are ready. The four new cases fail before the fix; the full Qdrant test module passes (47 tests) after it. Real in-memory sync/async roundtrips were also checked.

AI disclosure

AI-generated contribution: the investigation, reproduction, proposed implementation, tests, and this report were prepared and run with an AI coding assistant at the contributor's request. Please apply the required llm-generated label if my account cannot manage repository labels.