AI Agent + Vector Store Tool Generating Strange Responses and Missing Tool Information

Hi everyone,

I'm encountering a similar problem when utilizing the AI Agent with Llama 3.2 and Qdrant for vector storage. While Qdrant provides excellent responses, the AI Agent node seems to disregard them, opting instead to use its own chat model's output.

I'll be keeping an eye out for any developments regarding this issue.

For those encountering this problem, I experienced it too but managed to resolve it!

It seems that llama models are not optimal for embedding. Therefore, I've utilized the nomic-embed-text model for embedding. You can find it on Ollama.

My setup involves a single embedding node configured with the nomic model. This node connects to the Postgres PGVector Store for data insertion and also to the Postgres PGVector Store for retrieval.

I continue to use llama for the chat model.
This configuration works perfectly!

Hello everyone,

Just joining this discussion and wanted to ask, has anyone successfully used an AI agent node to answer questions based on documents stored in their vector database? I've found that the question and answer chain node works well, but I'm trying to have my AI agent node access a contacts document in my Pinecone database, retrieve contact information, and then send an email to a specified contact from my trigger prompt, utilizing another agent sub-workflow.

For some reason, I can't even get the agent to access the database.

image2862×686 91.2 KB

I am utilizing OpenAI for all my embeddings and chat nodes.

Found the solution for my situation, and this might just be basic troubleshooting:

1. I simply created a new, specific namespace in my Pinecone vector.
2. Updated the database retriever tool connected to the AI agent node with the new namespace.
3. It didn't query the namespace immediately after I set this up, so I wrote this post. Then, I went back to the workflow and reloaded the page.
4. It then worked, and I could query the Pinecone files using the chat trigger and AI agent.

Hope this helps!

I'm encountering a comparable problem: the AI agent isn't invoking the tools at all; it's providing all answers independently without querying the database.