Hi everyone. I've been trying to implement Relevance AI's 'parse email' tool into a workflow that handles customer support inquiries from a Gmail inbox. The goal is to extract text from attachments (like PDFs or DOCs) automatically. While it works perfectly about 60% of the time, the rest of the attempts fail silently or return empty content, specifically for Gmail attachments.
I've double-checked the obvious: the Gmail connector is set up correctly and has the necessary scopes (`gmail.readonly`, `gmail.modify`). The workflow triggers on new emails in the labeled inbox. The inconsistency is the frustrating part—same sender, similar attachment types, but unpredictable results.
Has anyone else run into this and found a reliable configuration? I'm particularly interested in:
- Whether you're using the native Gmail attachment data or a secondary step to process the file.
- Any specific pre-processing steps in the workflow before the 'parse email' tool runs.
- If there's a known issue with larger attachments or specific file types from Gmail's API.
I want to make sure we're comparing actual setups and results, not just guesses. Sharing your specific workflow structure would help us all nail down the culprit.
keep it evidence based
Ah, the classic "silently fails on Gmail attachments" dance. I've tripped over that one before. The issue you're describing often boils down to how the attachment data arrives from Gmail's API vs. what the parsing tool expects.
For us, the magic bullet was to never send the raw attachment data from the Gmail step directly into the parse tool. Instead, we added a step to first convert the attachment to base64 within the workflow before passing it along for parsing. It adds a small bit of overhead, but the consistency shot up to near 100%. The Gmail connector sometimes hands off the file in a way that the parser just shrugs at.
Also, watch out for scanned PDFs sent as images. If that's what your customers are sending, the basic parse tool will come back empty. You'd need an OCR step in there, which is a whole other can of worms.
It's always DNS.