# Discourse does not support Mistral structured content responses

**URL:** <https://meta.discourse.org/t/discourse-does-not-support-mistral-structured-content-responses/417710>\
**Category:** Bug\
**Tags:** ai\
**Created:** [18 november 2026 om 08:10 UTC](https://meta.discourse.org/t/discourse-does-not-support-mistral-structured-content-responses/417710 "2026-11-18T08:10:45Z")\
**Posts on this page:** 1\
**Page:** 1

## AI-generated summary

[RGJ](/t/-/417710/1) reports a bug in Discourse v2026.10.1 where the AI moderation feature crashes with a `TypeError: no implicit conversion of Hash into String` when using the `mistral-large-4` model. The error occurs during `LlmAgentTriage.handle` when attempting to moderate a post.

The root cause is identified as an incompatibility between Discourse's expected streaming format and Mistral Large 4's structured response format. While OpenAI-style APIs return `delta.content` as a simple string, Mistral Large 4 returns `delta.content` as an array of structured objects. These objects contain nested data, such as `thinking` blocks with type `"thinking"` and text blocks with type `"text"`. 

Discourse currently attempts to process this entire structure as a string, leading to the type conversion error. The reporter includes backtrace logs showing the issue originates in `discourse-ai/lib/completions/endpoints/base.rb` within the `streaming_partial_handler`. The proposed fix involves updating the Mistral endpoint to normalize the structured content by extracting chunks of type `"text"` for standard processing and handling `"thinking"` blocks separately, rather than treating the entire delta content as a string. This behavior aligns with the [Mistral reasoning documentation](https://docs.mistral.ai/inference/conversations/reasoning).

<div class="post-metadata">

**Author:** ![RGJ](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/rgj/32/523185_2.png) [@RGJ](https://meta.discourse.org/u/RGJ)\
**Post date:** [18 november 2026 om 08:10 UTC](https://meta.discourse.org/t/discourse-does-not-support-mistral-structured-content-responses/417710/1 "2026-11-18T08:10:45Z")

</div>

When moderating a post the following error occurs

`LlmAgentTriage.handle : TypeError : no implicit conversion of Hash into String`

Discourse v2026.10.1

Model: mistral-large-4

```plaintext
-[RECORD 2]------------+----------------------------------------------
id | 5
display_name | Mistral Large 4
name | mistral-large-4
provider | mistral
tokenizer | DiscourseAi::Tokenizer::MistralTokenizer
max_prompt_tokens | 8000
created_at | 2026-11-17 16:54:32.819593
updated_at | 2026-11-17 16:54:32.819593
url | https://api.eu.mistral.ai/v1/chat/completions
api_key |
user_id |
enabled_chat_bot | f
provider_params | {"disable_native_tools": null}
vision_enabled | f
input_cost |
cached_input_cost |
output_cost |
max_output_tokens | 2000
cache_write_cost | 0
allowed_attachment_types | {}
ai_secret_id | 1
vision_llm_model_id |

```

When moderating a post this happens

```plaintext

Error responding to: /t/testing-7-17-11-2026/2437/3 in LlmAgentTriage.handle : TypeError : no implicit conversion of Hash into String

Backtrace

/var/www/discourse/plugins/discourse-ai/lib/ai_bot/playground.rb:889:in 'block in DiscourseAi::AiBot::Playground#reply_to_unlocked'
/var/www/discourse/plugins/discourse-ai/lib/agents/bot.rb:461:in 'block in DiscourseAi::Agents::Bot#reply'
/var/www/discourse/plugins/discourse-ai/lib/completions/llm.rb:381:in 'block in DiscourseAi::Completions::Llm#generate'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:799:in 'block in DiscourseAi::Completions::Endpoints::Base#streaming_partial_handler'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:835:in 'block (3 levels) in DiscourseAi::Completions::Endpoints::Base#streaming_response'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:835:in 'Array#each'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:835:in 'block (2 levels) in DiscourseAi::Completions::Endpoints::Base#streaming_response'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:823:in 'Array#each'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:823:in 'block in DiscourseAi::Completions::Endpoints::Base#streaming_response'
net-protocol-0.4.0/lib/net/protocol.rb:604:in 'Net::ReadAdapter#call_block'
net-protocol-0.4.0/lib/net/protocol.rb:595:in 'Net::ReadAdapter#<<'
net-protocol-0.4.0/lib/net/protocol.rb:205:in 'Net::BufferedIO#read'
net-http-0.10.0/lib/net/http/response.rb:657:in 'Net::HTTPResponse#read_chunked'
net-http-0.10.0/lib/net/http/response.rb:616:in 'block in Net::HTTPResponse#read_body_0'
net-http-0.10.0/lib/net/http/response.rb:609:in 'Net::HTTPResponse#inflater'
net-http-0.10.0/lib/net/http/response.rb:614:in 'Net::HTTPResponse#read_body_0'
net-http-0.10.0/lib/net/http/response.rb:384:in 'Net::HTTPResponse#read_body'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:818:in 'DiscourseAi::Completions::Endpoints::Base#streaming_response'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:571:in 'block (3 levels) in DiscourseAi::Completions::Endpoints::Base#perform_completion_request_with_retries'
net-http-0.10.0/lib/net/http.rb:2520:in 'block in Net::HTTP#transport_request'
net-http-0.10.0/lib/net/http/response.rb:341:in 'Net::HTTPResponse#reading_body'
net-http-0.10.0/lib/net/http.rb:2517:in 'Net::HTTP#transport_request'
net-http-0.10.0/lib/net/http.rb:2463:in 'Net::HTTP#request'
rack-mini-profiler-5.0.0/lib/patches/net_patches.rb:19:in 'block in Net::HTTP#request_with_mini_profiler'
rack-mini-profiler-5.0.0/lib/mini_profiler/profiling_methods.rb:51:in 'Rack::MiniProfiler::ProfilingMethods#step'
rack-mini-profiler-5.0.0/lib/patches/net_patches.rb:18:in 'Net::HTTP#request_with_mini_profiler'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:522:in 'block (2 levels) in DiscourseAi::Completions::Endpoints::Base#perform_completion_request_with_retries'
net-http-0.10.0/lib/net/http.rb:1634:in 'Net::HTTP#start'
net-http-0.10.0/lib/net/http.rb:1068:in 'Net::HTTP.start'
/var/www/discourse/plugins/discourse-ai/lib/completions/endpoints/base.rb:503:in 'block in DiscourseAi::Completions::Endpoints::Base#perform_completion_request_with_retries'
:168:in 'Kernel#loop'

```

Logs have this

```plaintext

raw_response_payload | data: {"id":"bbda4b362069470e9c57d735628e0bbe","object":"chat.completion.chunk","created":1794935117,"model":"mistral-large-4","choices":[{"index":0,"delta":{"role":"assistant","content":""},"finish_reason":null}]} +
                     | +
                     | data: {"id":"bbda4b362069470e9c57d735628e0bbe","object":"chat.completion.chunk","created":1794935117,"model":"mistral-large-4","choices":[{"index":0,"delta":{"content":[{"type":"thinking","thinking":[{"type":"text","text":"The user is asking"}],"closed":true}]},"finish_reason":null}],"p":"abcdefghijklmnopqrstuvwxyz012"} 
                     

```

Where OpenAI’s streaming format uses a string for `delta.content`:

```json
{
  "choices": [{
    "delta": {
      "content": "The answer is..."
    }
  }]
}

```

Mistral Large 4 instead returns structured content blocks:

```json
"delta": {
  "content": [
    {
      "type": "thinking",
      "thinking": [
        {
          "type": "text",
          "text": "The user is asking"
        }
      ],
      "closed": true
    }
  ]
}

```

It can subsequently return `type: "text"` blocks within the same structure. See [Reasoning | Mistral Docs](https://docs.mistral.ai/inference/conversations/reasoning)

Discourse cannot handle this. It should normalize the structured content in the Mistral endpoint, extracting `type: "text"` chunks and handling `type: "thinking"` separately, instead of treating it like a string.
