how to get org-mode properties working with gptel?

[julianh-y]

I've been unable to use get gptel to use the values defined in the properties of an org heading as explained in extra-org-mode-conveniences:

Example org file:

#+title: gptel test


Some intro test goes here

* what is Emacs?
:PROPERTIES:
:GPTEL_MODEL: llama3:latest
:GPTEL_BACKEND: Ollama
:GPTEL_SYSTEM: You are a large language model living in Emacs and a helpful assistant. Respond with code only.
:GPTEL_TEMPERATURE: 0.1
:END:

when I execute gptel-send instead of my properties, this is what I see in *gptel-log* buffer:

{
  "gptel": "request body",
  "timestamp": "2024-07-01 00:01:02"
}
{
  "model": "gemma2:9b-instruct-q8_0",
  "messages": [
    {
      "role": "system",
      "content": "You are a large language model living in Emacs and a helpful assistant. Respond concisely."
    },
    {
      "role": "user",
      "content": "#+title: gptel test\n\n\nSome intro test goes here\n\n* what is Emacs?"
    }
  ],
  "stream": true,
  "options": {
    "num_ctx": 8192,
    "temperature": 1.0
  }
}

then the response is inserted in between the header line and the properties:

#+title: gptel test


Some intro test goes here

* what is Emacs?

Emacs is a free, open-source text editor with powerful extensibility features.
:PROPERTIES:
:GPTEL_MODEL: llama3:latest
:GPTEL_BACKEND: Ollama
:GPTEL_SYSTEM: You are a large language model living in Emacs and a helpful assistant. Respond with code only.
:GPTEL_TEMPERATURE: 0.1
:END:

I see the function is advised as expected and debugger trace looks like this:

Debugger entered--returning value: ("llama3:latest" "gemma2:9b-instruct-q8_0")
  seq-map(car (("llama3:latest" 0.1 nil) ("gemma2:9b-instruct-q8_0" 1.0 nil)))
* seq-mapn(#<subr F616e6f6e796d6f75732d6c616d626461_anonymous_lambda_3> ("You are a large language model living in Emacs and a helpful assistant. Respond with code only." #s(gptel-ollama :name "Ollama" :host "localhost:11434" :header nil :protocol "http" :stream t :endpoint "/api/chat" :key nil :models ("gemma2:9b-instruct-q8_0" "llama3:latest" "mistral:latest" "codellama:latest") :url "http://localhost:11434/api/chat" :curl-args nil) "llama3:latest" 0.1 nil) ("You are a large language model living in Emacs and a helpful assistant. Respond concisely." #s(gptel-ollama :name "Ollama" :host "localhost:11434" :header nil :protocol "http" :stream t :endpoint "/api/chat" :key nil :models ("gemma2:9b-instruct-q8_0" "llama3:latest" "mistral:latest" "codellama:latest") :url "http://localhost:11434/api/chat" :curl-args nil) "gemma2:9b-instruct-q8_0" 1.0 nil))
* gptel-org--send-with-props(#<subr gptel-send> nil)
* apply(gptel-org--send-with-props #<subr gptel-send> nil)
* #f(advice gptel-org--send-with-props :around #<subr gptel-send>)(nil)
* apply(#f(advice gptel-org--send-with-props :around #<subr gptel-send>) nil)
* gptel-send(nil)
  funcall-interactively(gptel-send nil)
  command-execute(gptel-send record)
  execute-extended-command(nil "gptel-send" nil)
  funcall-interactively(execute-extended-command nil "gptel-send" nil)
  command-execute(execute-extended-command)

I'm using GNU Emacs 31.0.50 with the following version of gptel

  gptel                          20240626.2246  installed             Interact with ChatGPT or other LLMs

[karthink]

Properties apply to the heading that they're in. Move your cursor to "inside" the heading, i.e. under the properties block and try it. If you want to set properties buffer-wide, call gptel-org-set-properties outside of any heading, or at the top of the buffer.

[julianh-y]

when I do that, the properties are sent as part of the content message instead of modifying the request parameters:

{
  "gptel": "request body",
  "timestamp": "2024-07-02 00:49:24"
}
{
  "model": "gemma2:9b-instruct-q8_0",
  "messages": [
    {
      "role": "system",
      "content": "You are a large language model living in Emacs and a helpful assistant. Respond concisely."
    },
    {
      "role": "user",
      "content": "#+title: gptel test\n\n\nSome intro test goes here\n\n* what is Emacs?\n:PROPERTIES:\n:GPTEL_MODEL: llama3:latest\n:GPTEL_BACKEND: Ollama\n:GPTEL_SYSTEM: You are a large language model living in Emacs and a helpful assistant. Respond with code only.\n:GPTEL_TEMPERATURE: 0.1\n:END:"
    }
  ],
  "stream": true,
  "options": {
    "num_ctx": 8192,
    "temperature": 1.0
  }
}

[karthink]

What version of Org mode are you using?

[julianh-y]

Org mode version 9.7.5 (release_9.7.5-9-ga091ca)

[karthink]

That should be fine. Have you set gptel-org-branching-context to t?

[julianh-y]

yes I'm setting it to t.

My gptel configuration, just in case I'm missing something or doing something wrong:

(use-package gptel
  :ensure t)

(setq gptel-temperature 0.0)

(setq gptel-model "gemma2:9b-instruct-q8_0"
      gptel-backend (gptel-make-ollama "Ollama"
		      :host "localhost:11434"
		      :stream t
		      :models '("gemma2:9b-instruct-q8_0" "llama3:latest" "mistral:latest" "codellama:latest")))

(setq gptel-default-mode 'org-mode)
(setq gptel-org-branching-context t)

(add-hook 'gptel-post-stream-hook 'fill-paragraph)

[karthink]

BTW to see what will be sent, you can use the dry run options in the transient menu. These are available once you've set gptel-log-level. You don't have to actually send the request and look at the log afterwards.