github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[PR #14288] server: structured outputs for thinking models in /api/generate #45851

New Issue
Open
opened 2026-04-25 01:28:09 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-25 01:28:09 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/14288
Author: @veeceey
Created: 2/17/2026
Status: 🔄 Open

Base: main ← Head: fix/issue-11691-structured-output

📝 Commits (2)

  • 63cffe7 server: add structured outputs support for thinking models in /api/generate
  • e2ca382 Merge branch 'main' into fix/issue-11691-structured-output

📊 Changes

3 files changed (+202 additions, -60 deletions)

View changed files

📝 middleware/openai_test.go (+26 -0)
📝 openai/openai_test.go (+40 -0)
📝 server/routes.go (+136 -60)

📄 Description

What

Adds the double-request structured outputs approach to the /api/generate endpoint, matching the existing implementation in /api/chat.

Why

The /api/chat endpoint got structured output support for thinking models (gpt-oss, etc.) in #12460, but /api/generate was left behind. This meant that any client using the generate endpoint with a format constraint on a thinking model would get either garbage output or empty responses — the format grammar was being applied during the thinking phase, corrupting the output.

Several users in #11691 confirmed that ollama-rs and other clients that go through /api/generate are still broken, even though /api/chat works.

How

Same pattern as the chat handler:

  1. When a format constraint is set and the model has thinking capability, the first completion request runs without the format constraint so the model can think freely.
  2. Once the parser detects that thinking is done and content is starting, the request is cancelled.
  3. The prompt is rebuilt with the thinking content appended as an assistant message (using chatPrompt), and for Harmony models the <|end|><|start|>assistant<|channel|>final<|message|> header is added.
  4. A second completion request runs with the format constraint active, producing valid structured output.

This only activates when:

  • A format constraint is set (req.Format != nil)
  • The model supports thinking (model.CapabilityThinking)
  • A builtin parser or thinking state parser is active
  • We're in the chat-like flow (have messages to rebuild the prompt)

For non-thinking models or raw mode, behavior is unchanged — the format is applied directly as before.

Tests

  • Added test for json_schema response format conversion in openai/openai_test.go
  • Added test for json_schema request handling in middleware/openai_test.go
  • All existing tests pass: go test ./server/ ./openai/ ./middleware/

Fixes #11691

🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/ollama/ollama/pull/14288 **Author:** [@veeceey](https://github.com/veeceey) **Created:** 2/17/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `fix/issue-11691-structured-output` --- ### 📝 Commits (2) - [`63cffe7`](https://github.com/ollama/ollama/commit/63cffe7aef274dd872f706594d58a3af34fe9f98) server: add structured outputs support for thinking models in /api/generate - [`e2ca382`](https://github.com/ollama/ollama/commit/e2ca3823678173056a1adabbc223ac975532e27a) Merge branch 'main' into fix/issue-11691-structured-output ### 📊 Changes **3 files changed** (+202 additions, -60 deletions) <details> <summary>View changed files</summary> 📝 `middleware/openai_test.go` (+26 -0) 📝 `openai/openai_test.go` (+40 -0) 📝 `server/routes.go` (+136 -60) </details> ### 📄 Description ## What Adds the double-request structured outputs approach to the `/api/generate` endpoint, matching the existing implementation in `/api/chat`. ## Why The `/api/chat` endpoint got structured output support for thinking models (gpt-oss, etc.) in #12460, but `/api/generate` was left behind. This meant that any client using the generate endpoint with a format constraint on a thinking model would get either garbage output or empty responses — the format grammar was being applied during the thinking phase, corrupting the output. Several users in #11691 confirmed that `ollama-rs` and other clients that go through `/api/generate` are still broken, even though `/api/chat` works. ## How Same pattern as the chat handler: 1. When a format constraint is set and the model has thinking capability, the first completion request runs **without** the format constraint so the model can think freely. 2. Once the parser detects that thinking is done and content is starting, the request is cancelled. 3. The prompt is rebuilt with the thinking content appended as an assistant message (using `chatPrompt`), and for Harmony models the `<|end|><|start|>assistant<|channel|>final<|message|>` header is added. 4. A second completion request runs **with** the format constraint active, producing valid structured output. This only activates when: - A format constraint is set (`req.Format != nil`) - The model supports thinking (`model.CapabilityThinking`) - A builtin parser or thinking state parser is active - We're in the chat-like flow (have messages to rebuild the prompt) For non-thinking models or raw mode, behavior is unchanged — the format is applied directly as before. ## Tests - Added test for `json_schema` response format conversion in `openai/openai_test.go` - Added test for `json_schema` request handling in `middleware/openai_test.go` - All existing tests pass: `go test ./server/ ./openai/ ./middleware/` Fixes #11691 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
GiteaMirror added the pull-request label 2026-04-25 01:28:09 -05:00
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 pull-request
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#45851
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?