github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 17:49:24 -05:00
Code Issues 14.6k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #11185] Ollama 0.9.2 never function_calls or tool_calls and fill content instead #69431

New Issue
Closed
opened 2026-05-04 18:04:17 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-05-04 18:04:17 -05:00
Owner
Copy Link

Originally created by @JpEncausse on GitHub (Jun 24, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/11185

What is the issue?

Ollama 0.9.2 NEVER call correctly Tools since 0.9.2 and there is no workaround. Please roll back Ollama !

  • ChatGPT advice to force post.function_call but it doesn't work better and LLM MUST choose tools
  • ChatGPT advice to declare a NOP fake tools ... WTF ?

Models like qwen3:32b REPLY tool calls in the content of the message and `qwen3:8b often hallucinate :

{
    "model": "qwen3:32b",
    "created_at": "2025-06-24T20:55:49.5546988Z",
    "message": {
        "role": "assistant",
        "content": "{\"name\": \"LLM_Tool_RAG\", \"arguments\": {\"term\": \"accidents, tribunal de Versailles\", \"vector\": \"accidents tribunal Versailles jugements\"}}"
    },
    "done_reason": "stop",
    "done": true,
    "total_duration": 16534413000,
    "load_duration": 21226300,
    "prompt_eval_count": 583,
    "prompt_eval_duration": 4321418400,
    "eval_count": 39,
    "eval_duration": 12187572000
}

Ollama MUST follow it's specifications and provides "tool_calls": [] and "done_reason": "tool",

{
  "model": "qwen3:32b",
  "created_at": "2025-06-24T20:55:49.5546988Z",
  "message": {
    "role": "assistant",
    "tool_calls": [
      {
        "name": "LLM_Tool_RAG",
        "arguments": {
          "term": "accidents, tribunal de Versailles",
          "vector": "accidents tribunal Versailles jugements"
        }
      }
    ]
  },
  "done_reason": "tool",
  "done": true,
  "total_duration": 16534413000,
  "load_duration": 21226300,
  "prompt_eval_count": 583,
  "prompt_eval_duration": 4321418400,
  "eval_count": 39,
  "eval_duration": 12187572000
}

The HTTP POST Request:


{
    "model": "qwen3:32b",
    "stream": false,
    "format": "json",
    "num_ctx": 8192,
    "temperature": 0.1,
    "num_predict": 4000,
    "tools": [
        {
            "type": "function",
            "function": {
                "name": "LLM_Tool_NOP",
                "description": "Dummy Tool that do nothing",
                "parameters": {
                    "type": "object",
                    "properties": {},
                    "required": []
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "LLM_Tool_RAG",
                "description": "Query a DATABASE about judgements issued by various French tribunals. ALL parameters are required.",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "term": {
                            "type": "string",
                            "description": "List of short and focused terms, comma separated."
                        },
                        "vector": {
                            "type": "string",
                            "description": "Search sentence build from the user query."
                        }
                    },
                    "required": [
                        "term",
                        "vector"
                    ]
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "LLM_Tool_HOME_Light",
                "description": "Switch On/Off Light. Change Colors",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "room": {
                            "type": "array",
                            "description": "Array of room to work with. \n        LIMITED VALUES: salon, bureau, chambre1, chambre2, entree, sallebain, couloir, escalier, cuisine, cave"
                        },
                        "dim": {
                            "type": "number",
                            "description": "Light luminosity between 0 (off) to 100 (on)"
                        },
                        "color": {
                            "type": "string",
                            "description": "A light color in hexa example: FFAACC"
                        }
                    },
                    "required": [
                        "room",
                        "dim"
                    ]
                }
            }
        },
        {
            "type": "function",
            "function": {
                "name": "LLM_Tool_INTERNAL_Image",
                "description": "You are Multimodal and can understand Image URL",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "image": {
                            "type": "string",
                            "description": "The image url"
                        }
                    },
                    "required": [
                        "image"
                    ]
                }
            }
        }
    ],
    "messages": [
        {
            "role": "system",
            "content": "Current date is: 24/06/2025 22:55:33\n\n# ROLE \n1. Analyze the user\u2019s message.\n2. Determine and Call function(s) tool(s) that are necessary to answer.\n3. After function or tool execution, respond with the final JSON:\n{\n  \"text\": \"<Markdown-formatted answer>\",\n  \"speech\": \"<Plain-text for TTS>\",\n}\n\n# FORMAT RULES\n- No additional text outside the JSON/blocs.\n- Follow the sequence strictly: analyze => tool_call => response.\n\nFunction or Tool MUST ALWAYS be declared in message.function_calls[] NEVER in message.content"
        },
        {
            "role": "user",
            "content": "Cherche les jugements d'accidents au tribunal de versailles. Ne demande pas de pr\u00e9cisions. Cherche dans le RAG."
        }
    ]
}

Relevant log output

OS

Windows

GPU

Nvidia

CPU

AMD

Ollama version

0.9.2

Originally created by @JpEncausse on GitHub (Jun 24, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/11185 ### What is the issue? Ollama 0.9.2 NEVER call correctly Tools since 0.9.2 and there is no workaround. Please roll back Ollama ! - ChatGPT advice to force `post.function_call` but it doesn't work better and LLM MUST choose tools - ChatGPT advice to declare a NOP fake tools ... WTF ? Models like `qwen3:32b` REPLY tool calls in the content of the message and `qwen3:8b often hallucinate : ``` { "model": "qwen3:32b", "created_at": "2025-06-24T20:55:49.5546988Z", "message": { "role": "assistant", "content": "{\"name\": \"LLM_Tool_RAG\", \"arguments\": {\"term\": \"accidents, tribunal de Versailles\", \"vector\": \"accidents tribunal Versailles jugements\"}}" }, "done_reason": "stop", "done": true, "total_duration": 16534413000, "load_duration": 21226300, "prompt_eval_count": 583, "prompt_eval_duration": 4321418400, "eval_count": 39, "eval_duration": 12187572000 } ``` Ollama MUST follow it's specifications and provides **"tool_calls": []** and **"done_reason": "tool",** ``` { "model": "qwen3:32b", "created_at": "2025-06-24T20:55:49.5546988Z", "message": { "role": "assistant", "tool_calls": [ { "name": "LLM_Tool_RAG", "arguments": { "term": "accidents, tribunal de Versailles", "vector": "accidents tribunal Versailles jugements" } } ] }, "done_reason": "tool", "done": true, "total_duration": 16534413000, "load_duration": 21226300, "prompt_eval_count": 583, "prompt_eval_duration": 4321418400, "eval_count": 39, "eval_duration": 12187572000 } ``` The HTTP POST Request: ``` { "model": "qwen3:32b", "stream": false, "format": "json", "num_ctx": 8192, "temperature": 0.1, "num_predict": 4000, "tools": [ { "type": "function", "function": { "name": "LLM_Tool_NOP", "description": "Dummy Tool that do nothing", "parameters": { "type": "object", "properties": {}, "required": [] } } }, { "type": "function", "function": { "name": "LLM_Tool_RAG", "description": "Query a DATABASE about judgements issued by various French tribunals. ALL parameters are required.", "parameters": { "type": "object", "properties": { "term": { "type": "string", "description": "List of short and focused terms, comma separated." }, "vector": { "type": "string", "description": "Search sentence build from the user query." } }, "required": [ "term", "vector" ] } } }, { "type": "function", "function": { "name": "LLM_Tool_HOME_Light", "description": "Switch On/Off Light. Change Colors", "parameters": { "type": "object", "properties": { "room": { "type": "array", "description": "Array of room to work with. \n LIMITED VALUES: salon, bureau, chambre1, chambre2, entree, sallebain, couloir, escalier, cuisine, cave" }, "dim": { "type": "number", "description": "Light luminosity between 0 (off) to 100 (on)" }, "color": { "type": "string", "description": "A light color in hexa example: FFAACC" } }, "required": [ "room", "dim" ] } } }, { "type": "function", "function": { "name": "LLM_Tool_INTERNAL_Image", "description": "You are Multimodal and can understand Image URL", "parameters": { "type": "object", "properties": { "image": { "type": "string", "description": "The image url" } }, "required": [ "image" ] } } } ], "messages": [ { "role": "system", "content": "Current date is: 24/06/2025 22:55:33\n\n# ROLE \n1. Analyze the user\u2019s message.\n2. Determine and Call function(s) tool(s) that are necessary to answer.\n3. After function or tool execution, respond with the final JSON:\n{\n \"text\": \"<Markdown-formatted answer>\",\n \"speech\": \"<Plain-text for TTS>\",\n}\n\n# FORMAT RULES\n- No additional text outside the JSON/blocs.\n- Follow the sequence strictly: analyze => tool_call => response.\n\nFunction or Tool MUST ALWAYS be declared in message.function_calls[] NEVER in message.content" }, { "role": "user", "content": "Cherche les jugements d'accidents au tribunal de versailles. Ne demande pas de pr\u00e9cisions. Cherche dans le RAG." } ] } ``` ### Relevant log output ```shell ``` ### OS Windows ### GPU Nvidia ### CPU AMD ### Ollama version 0.9.2
GiteaMirror added the bug label 2026-05-04 18:04:17 -05:00
GiteaMirror commented 2026-05-04 18:04:20 -05:00
Author
Owner
Copy Link

@jmorganca commented on GitHub (Jun 24, 2025):

Hi @JpEncausse sorry about the issue. "format": "json" will not work with tools as models require other control tokens to be output in order for successful tool calling. If you remove "format": "json" tool calling will work:

{"model":"qwen3:30b-a3b","created_at":"2025-06-24T21:18:31.216469Z","message":{"role":"assistant","content":"\u003cthink\u003e\nOkay, the user is asking to search for judgments related to accidents at the Versailles court using the RAG tool. Let me check the available functions. The LLM_Tool_RAG requires both 'term' and 'vector' parameters. The user mentioned \"jugements d'accidents\" and \"tribunal de Versailles\". So, I'll set 'term' to \"accidents, jugements\" and 'vector' to \"accidents at Versailles court\". I need to make sure both parameters are included. No other tools seem necessary here. Let me structure the tool call accordingly.\n\u003c/think\u003e\n\n","tool_calls":[{"function":{"name":"LLM_Tool_RAG","arguments":{"term":"accidents, jugements","vector":"accidents at Versailles court"}}}]},"done_reason":"stop","done":true,"total_duration":3292787792,"load_duration":23554542,"prompt_eval_count":583,"prompt_eval_duration":481477083,"eval_count":162,"eval_duration":2785346542}

In the future we will update the API to not allow format json with tool calling

<!-- gh-comment-id:3001890624 --> @jmorganca commented on GitHub (Jun 24, 2025): Hi @JpEncausse sorry about the issue. `"format": "json"` will not work with tools as models require other control tokens to be output in order for successful tool calling. If you remove `"format": "json"` tool calling will work: ``` {"model":"qwen3:30b-a3b","created_at":"2025-06-24T21:18:31.216469Z","message":{"role":"assistant","content":"\u003cthink\u003e\nOkay, the user is asking to search for judgments related to accidents at the Versailles court using the RAG tool. Let me check the available functions. The LLM_Tool_RAG requires both 'term' and 'vector' parameters. The user mentioned \"jugements d'accidents\" and \"tribunal de Versailles\". So, I'll set 'term' to \"accidents, jugements\" and 'vector' to \"accidents at Versailles court\". I need to make sure both parameters are included. No other tools seem necessary here. Let me structure the tool call accordingly.\n\u003c/think\u003e\n\n","tool_calls":[{"function":{"name":"LLM_Tool_RAG","arguments":{"term":"accidents, jugements","vector":"accidents at Versailles court"}}}]},"done_reason":"stop","done":true,"total_duration":3292787792,"load_duration":23554542,"prompt_eval_count":583,"prompt_eval_duration":481477083,"eval_count":162,"eval_duration":2785346542} ``` In the future we will update the API to not allow format json with tool calling
GiteaMirror commented 2026-05-04 18:04:21 -05:00
Author
Owner
Copy Link

@JpEncausse commented on GitHub (Jun 25, 2025):

OMG thanks for the answer !
You should really really details each parameters of the API.
Behaviors really change according to the usecase, model and/or the way to set them (HTTP, CLI, ...)

In my case, I ask for a structured JSON Answer that may requires calling tools before.

<!-- gh-comment-id:3003686085 --> @JpEncausse commented on GitHub (Jun 25, 2025): OMG thanks for the answer ! You should really really details each parameters of the API. Behaviors really change according to the usecase, model and/or the way to set them (HTTP, CLI, ...) In my case, I ask for a structured JSON Answer that may requires calling tools before.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label bug
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#69431
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?