I'm trying to find a nice way to integrate transparent tool-calling into an Open WebUI agent. Most tool-calling mechanisms are designed to hide the calls and their results from the user. This is problematic for OWUI because OWUI only stores the user-visible text of a conversation. To reconstruct the results of external actions, I either need to re-play the actions (hopefully yielding similar cached results) or create my own persistent storage system within OWUI. I'd rather just make everything visible in the public-facing text.
The idea I'm exploring here is to use the <details> syntax (ref) supported by many markdown renderers, including OWUI.
When asking the LLM to generate a reply, we'll use </summary> as a stop sequence. This way, a filter outlet can notice when a new tool call is attempted and insert the correct results (or a useful error message). When we splat in the results, we'll get the user interface to programmatically trigger a continuation of the response so that it gets a chance to react to what was looked up.
You have access to a wiki. Use syntax like this to retrieve pages:
⟨wiki path/to/file.md⟩
this is the contents of file.mdFiles:
- overview.md
- secret/hidden.md
You can discuss the content of any page with the user.
what's the wiki about?
⟨wiki overview.md⟩
we have a lot of secrets herecan you tell me about the hidden page?
It seems the "hidden" page is located in a restricted directory called "secret." I can only proceed with it if you ask explicitly. Should I open that page for you?
please do
⟨wiki secret/hidden.md⟩
buttslist all of the pages
The wiki contains the following pages:
- overview.md
- secret/hidden.md
Would you like to revisit or explore any of these further?