How to manage multi-part questions in RAG

1. Add a Function Node before your vector search
   This node will segment the query into distinct parts:

const input = $json.chatInput;
const parts = input.split(/and|&|then|also|,/i).map(p => p.trim()).filter(Boolean);
return parts.map(p => ({ question: p }));

2. Iterate through each segment using an Item Lists → Split Out node or a Split In Batches node.
3. Execute vector search for each sub-query. Link the output of the Function node to your Supabase vector search.
4. Gather and consolidate all responses using a Merge node (Combine mode: Append) or combine them manually in a final Function node.
5. Optional: Pass the consolidated response to a Chat Model node for a polished, natural-sounding answer:

This approach prevents vector store failures caused by improperly formatted or ambiguous multi-intent queries, ensuring the user receives a cohesive, multi-part answer.

Thanks! However, I don't believe this will work because you cannot insert a function node before the vector store. The vector store functions as a tool attached to the AI agent.

I've defined this within the AI Agent's prompt, and it appears to handle the initial part of the query adequately. Ideally, perhaps I should place something before the AI Agent to enable it to be called with multiple inputs.

{
  "userMessage": "{{(() => {
    const raw = $json.chatInput || $json.body?.question || $json?.input || '';
    if (typeof raw !== 'string') return 'Fallback query';
    const parts = raw.split(/[.?!]s+/).filter(p => p.length > 10);
    return parts[0] || raw;
  })()}}"
}

1. Divide the user’s query into segments before the AI Agent

• Insert a Function node directly after your trigger with the following code:

const raw = $json.chatInput || '';
const parts = raw
  .split(/[.?!]s+/)
  .map(p => p.trim())
  .filter(p => p);
return parts.map(p => ({ chatInput: p }));

• This will produce one item for each segment of the question, each within its own $json.chatInput.

2. Utilize a SplitInBatches node

• Link the Function node to a SplitInBatches node (set Batch Size to 1).
• This configuration ensures the AI Agent processes each segment individually.

3. Connect the AI Agent

• Feed the output from SplitInBatches → AI Agent (with your Supabase vector-store tool activated).
• The agent will subsequently invoke the vector store tool for each segment separately.

4. Consolidate the individual responses

• Following the AI Agent, add a Merge node:

  • Mode: Combine
  • Output Type: Append
  • This action gathers all responses from each segment into a single array.

5. (Optional) Structure the final reply

• Input the merged array into a Chat Model node with a prompt similar to this:

Here are the answers to each part of your question:
{{ $json.responses.join('n') }}

• Alternatively, use another Function node for concatenation.

This approach ensures that each sub-question is processed by the vector store tool independently, thereby preventing multi-intent errors.

Thanks, that guided me correctly. It appears to be functioning now.

Screenshot 2025-07-15 at 18.37.081564×502 143 KB

This thread was automatically closed 7 days following the last response. New replies are no longer permitted.