Ollama - Standardize meta

[saikanov]

Is your feature request related to a problem? Please describe.
No, its for enhancement for already working thing

Describe the solution you'd like
So i want to request if possible to make the output of Every Generator is Standardize to one Ouput type

I was working with OpenAI Generator all the time, and i want to use Ollama Generator now, the problem is the Ouput type is different and made an error on metadata output (response already good).

Thanks!

[anakin87]

Hello!
In general, we try to be consistent, but some components (including Ollama generators) have not been adapted.

Please explain better what the problem is... Is related to the usage key in meta?

[saikanov]

Hello! In general, we try to be consistent, but some components (including Ollama generators) have not been adapted.

Please explain better what the problem is... Is related to the usage key in meta?

Exactly! its the usage key in meta that raise me an error. well its just a minor error, but it could be good to be Standardize.

Thanks for your quick response!

[anakin87]

In short, we should do something like the following, both for OllamaGenerator and OllamaChatGenerator:

haystack-core-integrations/integrations/google_ai/src/haystack_integrations/components/generators/google_ai/chat/gemini.py

Lines 319 to 325 in 1ce24ce

	# adapt usage metadata to OpenAI format
	if usage_metadata := metadata.get("usage_metadata"):
	candidate_metadata["usage"] = {
	"prompt_tokens": usage_metadata["prompt_token_count"],
	"completion_tokens": usage_metadata["candidates_token_count"],
	"total_tokens": usage_metadata["total_token_count"],
	}

[lbux]

Ugh. I can take this over, but it won't be pretty. Unfortunately, switching to the Ollama Python library might have been a "regression" because we have effectively locked ourselves out of having 1:1 parity with OpenAI response objects. The best we would be able to do using Ollama's API is:

{
  "_role": "assistant",
  "_meta": {
    "usage": {
      "completion_tokens": 8,
      "prompt_tokens": 32,
      "total_tokens": null,                 # completion_tokens +  prompt_tokens might work ?
      "completion_tokens_details": null,
      "prompt_tokens_details": null
    },
    "model": "llama3.2:3b",
    "created_at": "2025-01-31T02:11:33.858103105Z",
    "done": true,
    "done_reason": "stop",                  # rename to finish_reason ?
    "total_duration": 279432961,
    "load_duration": 10745144,
    "prompt_eval_duration": 1000000,
    "eval_duration": 266000000
  },
  "_name": null,
  "_content": [
    {
      "text": "1 + 1 = 2"
    }
  ]
}

compared to OpenAI API's response:

{
  "_role": "assistant",
  "_meta": {
    "model": "llama3.2:3b",
    "index": 0,
    "finish_reason": "stop",
    "usage": {
      "completion_tokens": 8,
      "prompt_tokens": 32,
      "total_tokens": 40,
      "completion_tokens_details": null,
      "prompt_tokens_details": null
    }
  },
  "_name": null,
  "_content": [
    {
      "text": "1 + 1 = 2"
    }
  ]
}

But then again at this point we start diverging from the "Ollama API" and some users might expect the Ollama API.

Let me know what you think the best path forward is and I can make a PR @anakin87.