fix(openresponses): populate Content and accept bare {role,content} items (#10039) (#10040)

* fix(openresponses): populate Content and accept bare {role,content} items (#10039)

Fixes mudler/LocalAI#10039 — `/v1/responses` silently returned empty
output on any model whose YAML doesn't include a Go-side
`template.chat_message` block.

Three cooperating bugs:

* `convertORInputToMessages` populated only `StringContent` for string
  input and for the `input.Instructions` system message, leaving the
  `Content` (any) field nil.
* `TemplateMessages` gated all fallback content-rendering branches on
  `Content != nil && StringContent != ""` — but every branch in that
  function consumes `StringContent`, not `Content`. The `&&` silently
  dropped messages that had StringContent set and Content nil, producing
  an empty prompt that the 5× empty-retry guard then turned into a
  200 OK with `output: []`.
* The array-input branch of `convertORInputToMessages` dispatched on
  `itemMap["type"]` with no default, dropping bare `{role, content}`
  items emitted by the OpenAI Python SDK helper
  `client.responses.create(input=[{...}])`.

Fix:

* Set both `Content` and `StringContent` in the two openresponses
  message-construction sites that only set one.
* Treat a bare `{role, content}` item (no `type`) as
  `type: "message"` for OpenAI-SDK compatibility.
* Gate `TemplateMessages` fallback rendering on `StringContent != ""`,
  which is what every downstream branch in that function actually
  reads.

Regression test added to `evaluator_test.go` covering the fallback
path (no `ChatMessage` template) with a StringContent-only message,
both with and without a role mapping.

* test(openresponses): guard Content population and ToProto path (#10039)

Add regression tests for the two seams the original fix touched but
left uncovered:

* convertORInputToMessages must populate both Content and StringContent
  for plain string input and for bare {role, content} array items (the
  OpenAI SDK shape that omits the type discriminator). Both are
  functional reds against the pre-fix code.
* Messages.ToProto reads Content, not StringContent — this is the path
  UseTokenizerTemplate backends (imported GGUFs) take. The cases pin
  that contract so a future regression on the producer side is caught.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-7 [Claude Code]

---------

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
This commit is contained in:
Tai An
2026-05-28 00:21:48 -07:00
committed by GitHub
parent 7763fb23a3
commit 0fd666ee6e
5 changed files with 152 additions and 3 deletions

View File

@@ -95,7 +95,7 @@ func ResponsesEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, eval
// Add instructions as system message if provided
if input.Instructions != "" {
messages = append([]schema.Message{{Role: "system", StringContent: input.Instructions}}, messages...)
messages = append([]schema.Message{{Role: "system", Content: input.Instructions, StringContent: input.Instructions}}, messages...)
}
// Handle tools
@@ -299,7 +299,7 @@ func convertORInputToMessages(input any, cfg *config.ModelConfig) ([]schema.Mess
switch v := input.(type) {
case string:
// Simple string = user message
return []schema.Message{{Role: "user", StringContent: v}}, nil
return []schema.Message{{Role: "user", Content: v, StringContent: v}}, nil
case []any:
// Array of items
for _, itemRaw := range v {
@@ -309,6 +309,16 @@ func convertORInputToMessages(input any, cfg *config.ModelConfig) ([]schema.Mess
}
itemType, _ := itemMap["type"].(string)
// OpenAI SDK helpers (e.g. client.responses.create(input=[{"role":...,"content":...}]))
// send message items without a "type" discriminator. Treat a bare {role, content}
// object as type:"message" so the chat-completions and responses paths agree.
if itemType == "" {
if _, hasRole := itemMap["role"].(string); hasRole {
if _, hasContent := itemMap["content"]; hasContent {
itemType = "message"
}
}
}
switch itemType {
case "message":
msg, err := convertORMessageItem(itemMap, cfg)

View File

@@ -0,0 +1,62 @@
package openresponses
import (
"github.com/mudler/LocalAI/core/config"
. "github.com/onsi/ginkgo/v2"
. "github.com/onsi/gomega"
)
// Regression for mudler/LocalAI#10039. convertORInputToMessages must populate
// both Content and StringContent: the templating fallback path reads
// StringContent, while the UseTokenizerTemplate path serialises Content via
// Messages.ToProto(). Leaving Content nil produced an empty prompt on any model
// without a Go-side template.chat_message block (the default for imported GGUFs).
var _ = Describe("convertORInputToMessages", func() {
cfg := &config.ModelConfig{}
It("populates both Content and StringContent for plain string input", func() {
msgs, err := convertORInputToMessages("Hello", cfg)
Expect(err).NotTo(HaveOccurred())
Expect(msgs).To(HaveLen(1))
Expect(msgs[0].Role).To(Equal("user"))
Expect(msgs[0].StringContent).To(Equal("Hello"))
Expect(msgs[0].Content).To(Equal("Hello"))
})
It("accepts a bare {role, content} item without a type discriminator", func() {
// The OpenAI Python SDK helper client.responses.create(input=[{...}])
// sends message items with no "type" field. They must not be dropped.
input := []any{
map[string]any{"role": "user", "content": "Hi there"},
}
msgs, err := convertORInputToMessages(input, cfg)
Expect(err).NotTo(HaveOccurred())
Expect(msgs).To(HaveLen(1))
Expect(msgs[0].Role).To(Equal("user"))
Expect(msgs[0].StringContent).To(Equal("Hi there"))
Expect(msgs[0].Content).To(Equal("Hi there"))
})
It("still populates both fields for an explicit type:message item", func() {
input := []any{
map[string]any{"type": "message", "role": "user", "content": "Typed"},
}
msgs, err := convertORInputToMessages(input, cfg)
Expect(err).NotTo(HaveOccurred())
Expect(msgs).To(HaveLen(1))
Expect(msgs[0].StringContent).To(Equal("Typed"))
Expect(msgs[0].Content).To(Equal("Typed"))
})
It("does not treat a non-message item (no content key) as a message", func() {
// An item with neither a known type nor a {role, content} shape must
// keep falling through unchanged — no behaviour change for such inputs.
input := []any{
map[string]any{"role": "user"},
}
msgs, err := convertORInputToMessages(input, cfg)
Expect(err).NotTo(HaveOccurred())
Expect(msgs).To(BeEmpty())
})
})

View File

@@ -332,5 +332,41 @@ var _ = Describe("LLM tests", func() {
// Should only extract text parts
Expect(protoMessages[0].Content).To(Equal("Hello"))
})
// Regression for mudler/LocalAI#10039: ToProto is the path taken by
// UseTokenizerTemplate backends (e.g. imported GGUFs, where the backend
// applies the GGUF's jinja template to the raw messages). It reads
// Content, not StringContent — so a message that only populated
// StringContent (the shape /v1/responses produced before the fix)
// reached the backend with empty content. These two cases pin that
// contract: Content is authoritative, and producers must set it.
It("emits empty content when only StringContent is set (Content nil)", func() {
messages := Messages{
{
Role: "user",
StringContent: "Hello",
},
}
protoMessages := messages.ToProto()
Expect(protoMessages).To(HaveLen(1))
Expect(protoMessages[0].Content).To(BeEmpty())
})
It("carries Content through to proto regardless of StringContent", func() {
messages := Messages{
{
Role: "user",
Content: "Hello",
StringContent: "Hello",
},
}
protoMessages := messages.ToProto()
Expect(protoMessages).To(HaveLen(1))
Expect(protoMessages[0].Content).To(Equal("Hello"))
})
})
})

View File

@@ -111,7 +111,11 @@ func (e *Evaluator) TemplateMessages(input schema.OpenAIRequest, messages []sche
}
}
r := config.Roles[role]
contentExists := i.Content != nil && i.StringContent != ""
// Treat StringContent as the source of truth — every downstream fallback branch in this
// function reads StringContent, not Content. Gating on both with && silently drops
// messages that have StringContent set but Content nil (e.g. /v1/responses string-input
// before mudler/LocalAI#10039 fix).
contentExists := i.StringContent != ""
fcall := i.FunctionCall
if len(i.ToolCalls) > 0 {

View File

@@ -218,4 +218,41 @@ var _ = Describe("Templates", func() {
})
}
})
// Regression test for mudler/LocalAI#10039: when a model has no Go-side
// TemplateConfig.ChatMessage block (e.g. backends that rely on the GGUF's
// jinja template), TemplateMessages falls through to the role-prefix path.
// That path must still render messages whose StringContent is populated but
// Content (any) is nil — which is the shape /v1/responses produced before
// the fix to convertORInputToMessages.
Context("fallback path with StringContent-only message (no ChatMessage template)", func() {
var evaluator *Evaluator
BeforeEach(func() {
evaluator = NewEvaluator("")
})
It("renders the role prefix and content when only StringContent is set", func() {
cfg := &config.ModelConfig{
TemplateConfig: config.TemplateConfig{},
Roles: map[string]string{"user": "USER: "},
}
messages := []schema.Message{
{
Role: "user",
StringContent: "hello",
// Content intentionally left nil — reproduces /v1/responses string-input.
},
}
templated := evaluator.TemplateMessages(schema.OpenAIRequest{}, messages, cfg, []functions.Function{}, false)
Expect(templated).To(Equal("USER: hello"), templated)
})
It("renders content even with no role mapping", func() {
cfg := &config.ModelConfig{
TemplateConfig: config.TemplateConfig{},
}
messages := []schema.Message{
{Role: "user", StringContent: "hello"},
}
templated := evaluator.TemplateMessages(schema.OpenAIRequest{}, messages, cfg, []functions.Function{}, false)
Expect(templated).To(Equal("hello"), templated)
})
})
})