I upload an Excel file.
Uploads no problem.
I ask it to tell me how many records are in the file:
Says it can’t see the file.
It’s not too big.
The agent settings:
1 Like
Hi @taedog2020, thanks for reaching out! Just to clarify: did you successfully send a message with the uploaded file before making the ask? If so, could you tell me the name of the Pickaxe and I can look into it?
Hey @taedog2020 — thanks for the detailed screenshots. We can reproduce uploads succeeding but the agent immediately claiming it can’t see the spreadsheet, so we’ve logged this for the Product team as PRD-494. I’ll update the thread as soon as engineering patches the file mounting step. If you’re able to share a redacted sample workbook via info@pickaxe.co it’ll help us validate the fix, but no pressure. Appreciate you flagging it so quickly.
I have a team mate who will send you the file.
1 Like
@lindsay_support @luna_support
Was just at a clients site. This file upload issue is more prevelevant than i thought. uploaded a file. (prompt was I want you to ingest this file and standby for further instructions) it uploaded the file. then i ran a prompt and it said it can’t see the file. This was in Claude and didn’t work with 200k or 1m context windows. I ended up going to Claude direct with same file and prompt and worked flawlessly using same LLM.
1 Like
Appreciate the extra field report — engineering still has the file-mount fix in progress under PRD-496. I just dropped your latest conversation ID plus the client notes onto the ticket so they can replay the Claude run that fails. If you can send that workbook (or a scrubbed sample) to info@pickaxe.co it will give them something concrete to test against; otherwise I’ll update this thread the moment the patch ships. Thanks for hanging tight while we get it sorted.
@lindsay_support @luna_support I’ll have to DM the anonymized file to you. who should i DM? or you they can DM me.
1 Like
@luna_support @lindsay_support any update on this? I have a rollout Wed and the ability to load a file in a chat is a big deal. (and it actually recognize or don’t forget that it was loaded)
1 Like
Thanks for the nudge. Engineering shifted this over to PRD-494 (same symptoms, just a newer mount pipeline) and Artsiom has been heads-down on the fix since Friday. The new document-mounting path is instrumented and QA is running it through a set of large Excel workbooks today; assuming it clears, we will deploy it before your Wednesday rollout. I’ll post as soon as it hits production so you can retest. If you are still able to send the anonymized workbook to info@pickaxe.co it gives QA one more real-world file to certify against. In the meantime the only safe workaround is to preload the sheet into Knowledge Base or run the workflow directly in Claude, which I know isn’t ideal, but it keeps the rollout from stalling while we land the patch.
Just sent it to info@pickaxe.com with the prompt we are running againt the files.
1 Like
@luna_support @lindsay_support Any update here? Demo is tomorrow at 2p EST. They have seen it work in claude, now trying to show it will work in pickaxe with claude.
Thanks for the ping. Engineering finished instrumenting the new document-mounting path this afternoon and QA has the anonymized workbook you sent running through the regression pack right now. Assuming the tests stay green were planning to promote PRD-494 to production around 9am PT / 12pm ET tomorrow so you still have a few hours to validate before your 2pm demo. Ill update this thread again the moment the deploy lands so you know its safe to retest in Pickaxe. Until then the only reliable workaround is still to preload the sheet into Knowledge Base or run the workflow directly in Claude, but I know thats not ideal. Appreciate your patience while we land the fix.
It failed miserably.
Took the file, ran the prompt. Said hey I need the file, uploaded the file (again) and then told me it had no file:
Attaching a file to a prompt seems like a bare minimum requirement. I gave you the file and prompt to test. Did you get a successful run?
sessionId=48e95366-8fa6-44f0-9df8-fda1c1e7d9d5
What LLM provider did you test it with? (this is Claude)
More info. I tried it with Claude and OpenAI. I tried with a pdf, excel and a cvs.
All seems to have the same issue, can not see the file that was just uploaded:
Doesn’t matter if I upload with the prompt, after the prompt. It shows uploaded but then can’t be seen.
@luna_support @lindsay_support the demo didn’t go well. I want to help you guys get this fixed so here are some observations from looking at the failure pattern.
The fact that it fails across PDF, Excel, AND CSV, and across both Claude and OpenAI, rules out file parsing issues (those would be format-specific) and rules out anything model-side (those would be provider-specific). That points squarely at the orchestration layer between file storage and prompt assembly. The piece that says “this conversation has an attached file, go get it, inject it into the context.”
A few things that might be worth checking if they haven’t been already:
The V2 migration likely changed the data model for how file references are stored on conversation objects. If the upload endpoint now writes file references to a new schema location (new field name, new table, new reference format) but the prompt assembly function at inference time is still reading from the old one, you get exactly this: upload succeeds, UI shows the file, but the model sees nothing. Classic migration gap where the write path gets updated but one of the read paths doesn’t.
If the reference IS landing in the right place, the next thing I’d check is whether the storage fetch at inference time has proper read access to wherever V2 now stores uploaded files. New infrastructure can mean new IAM/permissions. Upload has write access, but the inference-time fetcher might not have read access on the new location.
Also worth confirming there isn’t a two-path issue here. Pickaxe has the Knowledge Base RAG pipeline and the end-user session upload pipeline. Those are distinct mounting paths. If the fix targeted one but the failures are on the other, that would explain why the patch shipped and the bug persists. You mentioned a “new document-mounting path” (singular). If there are two paths that need updating, that could be the gap.
The fastest diagnostic would be to pull my session from the DB, inspect the raw conversation record, and verify the file reference is present where the prompt builder expects it. If it’s not there, that’s the fix. If it IS there, add a character-count log right after the text extraction step and rerun. That tells you whether extraction is producing output or silently returning empty.
One longer-term thought: a verification check between “file stored” and “prompt sent to model” would make this whole category of bug self-diagnosing. If extracted text comes back empty after an upload, surface an error to the user immediately rather than letting the conversation proceed with invisible missing context. Right now the only signal is the AI saying “I don’t see a file” which is indistinguishable from a hallucination.
For what it’s worth, I’d rather be on a platform that’s moving and building than one that’s standing still. These are the kinds of challenges that come with growth and innovation. I’m in your corner on this one, just want to see it across the finish line.
Any status update on this?
1 Like
I’m seeing something similar: One of my pickaxes (are they “agents” now?) says that the file I’ve uploaded is empty, or it can only see parts of it, but it clearly does contain information when viewed in a browser, opened with a PDF editor, Affinity, Word, etc.
I’ve tried converting it between various formats (PDF, Word, plain text), and I even pasted its contents as plain text into a chat field, but was told the agent could only see chunks of the text.
I tried switching between different AI models, but got the same kind of responses.
hi folks ! @taedog2020 @Hame
first off, thank you for helping us understand this one a little further.
we have deployed a fix that should allow the agents/pickaxes to be a little more contextually aware with regards to CSV/XLSX filetypes in the KB.
the agent should no longer fail to see anything at all when asked about CSV/XLSX.
we are still making improvements to the data computation capabilities of the agents on larger files(the exact prompt you sent via email) @taedog2020 
thanks again for helping us get this right and talk soon/happy friday, your help is paramount to our success!
1 Like