mem0.add() does not store embeddings in Qdrant when using Ollama (mxbai-embed-large); returns {'results': []}

[ankurphani]

🐛 Describe the bug

This is probably a bug. I observed that the embeddings are not stored in quadrant store using .add().

It does not return any error message either.

Here is my config.

from mem0 import Memory

config = {
    "vector_store": {
        "provider": "qdrant",
        "config": {
            "collection_name": "test",
            "host": "localhost",
            "port": 6333,
            "embedding_model_dims": 1024,
        }
    },
    "llm": {
        "provider": "ollama",
        "config": {
            "model": "llama3.1:latest",
            "temperature": 0,
            "max_tokens": 2000,
            "ollama_base_url": "http://localhost:11434",
        },
    },
    "embedder": {
        "provider": "ollama",
        "config": {
            "model": "mxbai-embed-large",
            "ollama_base_url": "http://localhost:11434",

        },
    }
}

# Initialize Memory with the configuration
try:
    m = Memory.from_config(config)
except Exception as e:
    print(f"Error initializing Memory: {e}")
    exit(1)

# Add a memory
messages = [
    {"role": "user", "content": "I'm planning to watch a movie tonight. Any recommendations?"},
    {"role": "assistant", "content": "How about a thriller movies? They can be quite engaging."},
    {"role": "user", "content": "I’m not a big fan of thriller movies but I love sci-fi movies."},
    {"role": "assistant", "content": "Got it! I'll avoid thriller recommendations and suggest sci-fi movies in the future."}
]
try:
    m.add(messages, user_id="john")

except Exception as e:
    print(f"Error adding memory: {e}")
    exit(1)

# Retrieve memories
try:
    related_memories = m.search(query="What do you know about me?", user_id="john")
    print(f"related_memories: {related_memories}")
except Exception as e:
    print(f"Error retrieving memories: {e}")
    exit(1)

The above code returns

{'results': []}

I manually upserted the data in qdrant store using their REST API and it worked fine.

I have verified that the embedding model dimensions are right. i.e. 1024

resp = requests.post("http://localhost:11434/api/embeddings", json={
    "model": "mxbai-embed-large",
    "prompt": "dimension check"
})
vec = resp.json()["embedding"]
print(len(vec))  # this prints 1024

Can anyone confirm having this same issue using Ollama and Qdrant or there is some configuration issue from my side.

[ankurphani]

While debugging this, I figured if I switch my llm config from ollama to gemini, it works just fine.

This is my updated config.

os.environ["GOOGLE_API_KEY"] = "key"

config = {
    "vector_store": {
        "provider": "qdrant",
        "config": {
            "collection_name": "test02",
            "host": "localhost",
            "port": 6333,
            "embedding_model_dims": 1024,
        }
    },
    "llm": {
        "provider": "gemini",
        "config": {
            "model": "models/gemini-2.0-flash",
            "temperature": 0,
        },
    },
    "embedder": {
        "provider": "ollama",
        "config": {
            "model": "mxbai-embed-large",
            "ollama_base_url": "http://localhost:11434",

        },
    }
}

I am able to re-create it if I switch back to ollama.

[rkhatta1]

Interesting. The simple answer is: skill issue on llama's part :P

Jokes apart, here's what I suggest:

use m.add(messages, user_id="john", infer=False)

What I think is the issue:

• By default, infer is set to True, and this essentially makes mem0 "smart" (in this instance), where it uses an LLM to extract key "facts" from the conversation before embedding them. It analyzes the conversation, extracts key facts, and can even update or delete old memories if new information contradicts them.
• Now, the issue with llama3.1 vs gemini is: gemini is just plain smarter, larger and running on better infra, which translates to --> it consistently returns a valid json, which is in turn correctly interpreted and handled by the _add_to_vector_store() function. Llama3.1:latest on the other hand, in your case, is a relatively tiny model, hence, it is inconsistent with the json it outputs, which is the json the said function demands, if infer=True.
• Check this main.py: The try...except block on line 363 --> It sets the new_retrieved_facts as an empty dict, in case of an exception (an invalid json). This empty dict also breaks the rest of the code and the function ends up adding nothing to the memory. Hence, your current code returns {'results': []}.

Feel free to correct me if I misinterpreted something here.

[parshvadaftari]

Hey @phaniankur thank you for sharing the issue faced, the error faced is caused by the hallucinations by the LLM, for not returning the JSON response. Also in some cases this happens due to GPU limitations

@rkhatta1 Thanks a lot for helping, yes you're correct.

[parshvadaftari]

Closing this issue as it's fixed.