5 replies

Binary data injected into context when uploading pdf through telegram

✅Solved

I am having an issue with attaching pdf through telegram. i already have a scipt which when given a pdf file can extract the content of it and summarise it. This script basically has a UI (mission control) through which I can upload the pdf to it. But I want to interact with openclaw through telegram interface. The problem is as soon I attch the pdf, it the model (GLM 4.7) receives the file path (/home/ubuntu/.openclaw/media/inbound/sample.pdf) and also the mime type. SOmetimes the mime type is incorrectly identified as text for pdf documents. Here is an excerpt from one of the sessiosn

{"type":"message","id":"7f02c908","parentId":"36cb46e8","timestamp":"2026-02-06T10:38:09.850Z","message":{"role":"user","content":[{"type":"text","text":"[media attached: /home/ubuntu/.openclaw/media/inbound/file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf (application/pdf) | /home/ubuntu/.openclaw/media/inbound/file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf]\nTo send an image back, prefer the message tool (media/path/filePath). If you must inline, use MEDIA:https://example.com/image.jpg (spaces ok, quote if needed) or a safe relative path like MEDIA:./image.jpg. Avoid absolute paths (MEDIA:/...) and ~ paths — they are blocked for security. Keep caption in the text body.\nSystem: [2026-02-06 10:35:37 UTC] Refresh tasks from Google Tasks (auto-sync every 5 minutes)\n\n[Telegram John (@SJohn_doe) id:12240321 +2m 2026-02-06 10:38 UTC] <media:document> <file name=\"file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf\" mime=\"text/plain\"> %PDF-1.7 %���� 1 0 obj << /Metadata 3 0 R /Names 4 0 R /OpenAction 5 0 R /Outlines 6 0 R....... continued binary data

{"type":"message","id":"7f02c908","parentId":"36cb46e8","timestamp":"2026-02-06T10:38:09.850Z","message":{"role":"user","content":[{"type":"text","text":"[media attached: /home/ubuntu/.openclaw/media/inbound/file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf (application/pdf) | /home/ubuntu/.openclaw/media/inbound/file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf]\nTo send an image back, prefer the message tool (media/path/filePath). If you must inline, use MEDIA:https://example.com/image.jpg (spaces ok, quote if needed) or a safe relative path like MEDIA:./image.jpg. Avoid absolute paths (MEDIA:/...) and ~ paths — they are blocked for security. Keep caption in the text body.\nSystem: [2026-02-06 10:35:37 UTC] Refresh tasks from Google Tasks (auto-sync every 5 minutes)\n\n[Telegram John (@SJohn_doe) id:12240321 +2m 2026-02-06 10:38 UTC] <media:document> <file name=\"file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf\" mime=\"text/plain\"> %PDF-1.7 %���� 1 0 obj << /Metadata 3 0 R /Names 4 0 R /OpenAction 5 0 R /Outlines 6 0 R....... continued binary data

As a result the LLm model receives: "[media attached: PATH] + <binary data>". LLM sees gibberish, hallucinates a response. As a result LLM is not able to apply the same extreat_and_summarise script ot this pdf file. I want openclaw to only forward the file_path and not any extracted info in the context

Solution

Result: the model should still see the normal

[media attached: /home/...pdf (...)]

[media attached: /home/...pdf (...)]

line (so you can pass that path into your

extract_and_summarise

extract_and_summarise

flow), but it won’t get the decoded file bytes.

Docs where this config lives (it’s documented under the OpenResponses endpoint, but currently reused by the general attachment extractor too):
https://docs.openclaw.ai/gateway/openresponses-http-api#file--image-limits-config

If you want: help us fix the underlying PDF MIME mis-detect

If you paste:
-

openclaw --version

openclaw --version

- and (if possible) a short log snippet around where the attachment is processed

…I can point you to the exact bug/PR shape (it likely needs “don’t override

application/pdf

application/pdf

text/plain

text/plain

just because the first 4KB are printable”).

If this workaround fixes it for you, please right-click → Apps → Mark Solution so it’s indexed on Answer Overflow.

Jump to solution

FotC�

Friends of the Crustacean 🦞🤝•6d ago•

5 replies

Clawdbot

Binary data injected into context when uploading pdf through telegram

✅Solved

{"type":"message","id":"7f02c908","parentId":"36cb46e8","timestamp":"2026-02-06T10:38:09.850Z","message":{"role":"user","content":[{"type":"text","text":"[media attached: /home/ubuntu/.openclaw/media/inbound/file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf (application/pdf) | /home/ubuntu/.openclaw/media/inbound/file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf]\nTo send an image back, prefer the message tool (media/path/filePath). If you must inline, use MEDIA:https://example.com/image.jpg (spaces ok, quote if needed) or a safe relative path like MEDIA:./image.jpg. Avoid absolute paths (MEDIA:/...) and ~ paths — they are blocked for security. Keep caption in the text body.\nSystem: [2026-02-06 10:35:37 UTC] Refresh tasks from Google Tasks (auto-sync every 5 minutes)\n\n[Telegram John (@SJohn_doe) id:12240321 +2m 2026-02-06 10:38 UTC] <media:document> <file name=\"file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf\" mime=\"text/plain\"> %PDF-1.7 %���� 1 0 obj << /Metadata 3 0 R /Names 4 0 R /OpenAction 5 0 R /Outlines 6 0 R....... continued binary data

{"type":"message","id":"7f02c908","parentId":"36cb46e8","timestamp":"2026-02-06T10:38:09.850Z","message":{"role":"user","content":[{"type":"text","text":"[media attached: /home/ubuntu/.openclaw/media/inbound/file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf (application/pdf) | /home/ubuntu/.openclaw/media/inbound/file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf]\nTo send an image back, prefer the message tool (media/path/filePath). If you must inline, use MEDIA:https://example.com/image.jpg (spaces ok, quote if needed) or a safe relative path like MEDIA:./image.jpg. Avoid absolute paths (MEDIA:/...) and ~ paths — they are blocked for security. Keep caption in the text body.\nSystem: [2026-02-06 10:35:37 UTC] Refresh tasks from Google Tasks (auto-sync every 5 minutes)\n\n[Telegram John (@SJohn_doe) id:12240321 +2m 2026-02-06 10:38 UTC] <media:document> <file name=\"file_5---ee578df2-efd9-432d-b482-96fa886a379a.pdf\" mime=\"text/plain\"> %PDF-1.7 %���� 1 0 obj << /Metadata 3 0 R /Names 4 0 R /OpenAction 5 0 R /Outlines 6 0 R....... continued binary data

Solution

Result: the model should still see the normal

[media attached: /home/...pdf (...)]

[media attached: /home/...pdf (...)]

line (so you can pass that path into your

extract_and_summarise

extract_and_summarise

If you want: help us fix the underlying PDF MIME mis-detect

If you paste:
-

openclaw --version

openclaw --version

- and (if possible) a short log snippet around where the attachment is processed

…I can point you to the exact bug/PR shape (it likely needs “don’t override

application/pdf

application/pdf

text/plain

text/plain

just because the first 4KB are printable”).

If this workaround fixes it for you, please right-click → Apps → Mark Solution so it’s indexed on Answer Overflow.

Jump to solution

Binary data injected into context when uploading pdf through telegram

If you want: help us fix the underlying PDF MIME mis-detect

Binary data injected into context when uploading pdf through telegram

If you want: help us fix the underlying PDF MIME mis-detect

Similar Threads

Similar Threads

Similar Threads