Workflow Setup RAG Sales Agent - losing metadata from supabase vector store

[MartinSand]

Describe the problem/error/question

Hi,
i am relatively new to n8n and have no coding experience.
I try to build an ai sales consultant. The setup is attached.

Goal:

1. User types in chat message
2. RAG AI Agent finds the right answer to his questions, reply and asks the user if he likes to get a product recommendation (data is stored in supabase vector store; table documents - vector and 1-n chunks per question) –> Already Working
3. RAG AI Agent finds the right products for recommendation (data is stored in supabase vector store; table recommendations)
   The connection between the two tables is the question_key [frage_key)

So far so good.

How could i solve this? Is my setup maybe not good

Idea would be:
Ask the user if he likes to have a recommendation, then show him the available products.
The thing is how to get the right question_key.

What is the error message (if any)?

The challenge:

• The superbase vector store returns the metadata (i need the frage_key, which is the key to find the relevant products (recommendations) in the table recommendations) → the vector store question and answer tool only has the response as an output

• For the recommendations tool to work properly i need to have the [frage_key]

  

Please share your workflow

(Select the nodes on your canvas and use the keyboard shortcuts CMD+C/CTRL+C and CMD+V/CTRL+V to copy and paste the workflow.)

Share the output returned by the last node

Information on your n8n setup

• n8n version: 1.109.2
• Database (default: SQLite): supabase
• n8n EXECUTIONS_PROCESS setting (default: own, main):
• Running n8n via (Docker, npm, n8n cloud, desktop app): n8n cloud
• Operating system:

[krisn0x]

Hey, you can create a structured output node and also instruct the AI in its system message to always output the frage_key. Something like this:

[MartinSand]

Hi @krisn0x ,
thank you very much for your answer, but i tried and it is not working.
The Vector Store Question and Answer Tool only gives the response, no frage_key is sent to the RAG AI Agent.

[krisn0x]

Hey, you need to connect the Structured Output node. First, enable Require specific output in the AI node and connect them, then try again.

[MartinSand]

[MartinSand]

there is still the error. connected it before, but there are issues with the format:

Model output doesn’t fit required format

To continue the execution when this happens, change the ‘On Error’ parameter in the root node’s settings

• changed system prompt in RAG AI Agent
• used structured output parser

[krisn0x]

Hey, this is now either a prompt issue or issue with the agent finding the frage_key. I think the first one’s more likely.

Can you change the structure in the schema to add descriptions:

{
    "type":"object",
	"response": {"description": "Your description here",
                 "type":"string"},
	"frage_key": {"description": "Your description here",
                 "type":"string"},
}

And in your prompt you must be VERY specific:

• use words like ALWAYS output the frage_key associated with the output info
• tell it where to get the frage_key from (use the name of the tool)
• don’t define the schema in the prompt, it can confuse it as it’s already in the Output Parser
• disable the output parser for testing, so you can see how the output changes when you adjust the prompt

This will require prompt tuning above all else

[MartinSand]

Hi,
so i tested your suggestions:

• Structured Output Parser connected: no result → same error like before
• Structured Output Parser not connected: answer is right, still only output → it does not contain the frage_key information
  • my research last week said it is due to the “Vector Store Question and Answer Tool”
  • can you maybe think of another solution to get the frage_key to use then the recommendations tool?

[MartinSand]

results from supabase vector store:

FragenundAntworten Tool:

[krisn0x]

Ok, so, I can’t really test this for you as the setup is not reproducible without your vector store. What I would do next is this:

• Test with extremely basic prompt. Something like

Consult the <vector_tool_name> and query for X. The JSON you will receive from the <vector_tool_name> also contains a key called "frage_key". Output only the "farge_key" value.

• Try other models like Gemini and see if they accomplish the task. Your primary goal is to find out if any model can output the farge_key and nothing else matters (excuse the pun) for now.
• Once it can output the farge_key (if possible at all), get your prompt and pass it through another AI to refine it. Explain your issues and needs, so it can design it for you. I translated it to check it and see a lack of clarity around certain tools and sentences. Of course, keep the part that successfully printed the farge_key at the top and unchanged.

These are the tips I can think of right now. It’s a bit of a pain to fiddle with prompts on AI nodes, I know. Crossing fingers it takes less time rather than more!