[Bug]: Ollama provider returns empty model list after upgrading to v0.26.0 — Go rewrite fails to parse /api/tags response (original) (raw)

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (Language Policy).
• Non-english title submitions will be closed directly ( 非英文标题的提交将会被直接关闭 ) (Language Policy).
• Please do not modify this template :) and fill in all the required fields.

RAGFlow workspace code commit ID

v0.26.0

RAGFlow image version

v0.26.0

Other environment information

| RAGFlow | v0.26.0 (docker image) | | Ollama | v0.5.0 | | OS | CentOS 7 | | GPU | DCU Z100L-E4x16 32GB (x2) | | Deployment | Docker containerized |

Actual behavior

After upgrading RAGFlow from v0.25.5 to v0.26.0, the Ollama provider's "auto-populated model list" feature returns an empty list, making it impossible to add Ollama models through the UI. The connection test passes, but no models are discovered.

Actual behavior

• UI Error: 102 No valid models found for provider 'Ollama'
• UI Error: Your API key is invalid (Ollama does not require API Key)
• Backend: GET /api/v1/providers/Ollama/instances returns HTTP 200 with ~21 bytes (empty array)
• Backend: GET /api/v1/providers/Ollama/models returns HTTP 200 with ~63 bytes (nearly empty)

RAGFlow Backend Logs

Connection test passes

POST /api/v1/providers/Ollama/connection 1.1 200 69 6069

But model list is empty

GET /api/v1/providers/Ollama/instances 1.1 200 21 7513
GET /api/v1/providers/Ollama/models 1.1 200 63 5756

OpenAI-API-Compatible fallback also fails

Fail to access chat model for provider=OpenAI-API-Compatible model=qwen2.5:32b-instruct-q4_K_M
openai.NotFoundError: 404 page not found

Ollama /api/tags Response

The Ollama server is healthy and returns valid model data:

{ "models": [ { "name": "qwen2.5-coder:32b-8k", "model": "qwen2.5-coder:32b-8k", "modified_at": "2026-05-26T14:13:16.512592158+08:00", "size": 19835149999, "digest": "0b853175d5b9f40a6513ef8751c509744a7bf4230720e8399e465d543d3890", "details": { "parent_model": "", "format": "gguf", "family": "qwen2", "families": ["qwen2"], "parameter_size": "32.8B", "quantization_level": "Q4_K_M" } }, { "name": "qwen2.5-coder:14b", "model": "qwen2.5-coder:14b", "modified_at": "2026-05-21T15:34:27.161982701+08:00", "size": 8983124336, "digest": "0e58897f747e246e970bccf6da85d2f1123de2e3468730a010e75986716849", "details": { "parent_model": "", "format": "gguf", "family": "qwen2", "families": ["qwen2"], "parameter_size": "14.8B", "quantization_level": "Q4_K_M" } }, { "name": "qwen2.5-coder:32b", "model": "qwen2.5-coder:32b", "modified_at": "2026-05-21T02:45:21.550971622+08:00", "size": 19835149999, "digest": "b95685f0b4a7e791142936e0177f69200c5a706126328950b3926778492e4f1", "details": { "parent_model": "", "format": "gguf", "family": "qwen2", "families": ["qwen2"], "parameter_size": "32.8B", "quantization_level": "Q4_K_M" } }, { "name": "qwen2.5-32b-instruct-q4_K_M", "model": "qwen2.5-32b-instruct-q4_K_M", "modified_at": "2026-05-21T00:44:17.308569422+08:00", "size": 19315434959, "digest": "9f163a1309afea09d9a956c6a855ecc99115a6d596fae201a5457a035bd4365", "details": { "parent_model": "", "format": "gguf", "family": "qwen2", "families": ["qwen2"], "parameter_size": "32.8B", "quantization_level": "Q4_K_M" } } ] }

Expected behavior

The model list should populate with the 4 available Ollama models, as it did in v0.25.6.

Steps to reproduce

```Markdown

1. Deploy Ollama v0.5.0 (DCU Z100L environment, cannot upgrade Ollama)
2. Pull models: qwen2.5-coder:32b, qwen2.5-coder:14b, qwen2.5-coder:32b-8k, qwen2.5-32b-instruct-q4_K_M
3. Upgrade RAGFlow from v0.25.5 to v0.26.0
4. Go to Settings > Model Providers > Ollama > Add Model
5. Enter Base URL: http://10.206.254.5:11434
6. Click "Model List" or "Verify"
7. Observe: "102 No valid models found for provider 'Ollama'" and "Your API key is invalid"

Additional information

RAGFlow v0.26.0 rewrote the Ollama provider from Python to Go (#14580, #15213). The new Go implementation appears to have stricter parsing logic for the /api/tags response. Possible causes:
Family whitelist filtering: Go code may only recognize specific details.family values (e.g., "llama", "mistral") and filter out "qwen2"
Parameter_size format parsing: The "32.8B" format may not match Go's expected regex/pattern
Empty parent_model handling: parent_model: "" may cause JSON unmarshaling to fail in Go
API Key validation regression: Go code may incorrectly enforce non-empty API Key for Ollama

I cannot upgrade Ollama because:
Hardware: DCU Z100L (AMD ROCm-like architecture, not CUDA)
Ollama v0.5.0 is the last version with DCU support in my environment
Upgrading Ollama would break the entire inference pipeline