Skip to content

Agent Zero AI framework

Prevent chat parameters from leaking into embeddings

Alessandro71b9759d5117
diff --git a/AGENTS.md b/AGENTS.md
index 9a790fff9..95090ed9f 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -32,6 +32,7 @@
- Message-loop completion flows through a response tool with `break_loop`; plain or malformed Chat Completions text enters repair, and native Responses output text is normalized through the same response-tool path.
- Reuse the startup-preloaded local embedding model for matching runtime configurations; wrappers retain their own rate-limit configuration while sharing the underlying inference model.
- Embedding wrappers expose batch-shaped `embed(inputs)` for provider-ready inputs; keep `embed_documents` and `embed_query` as LangChain compatibility adapters.
+- Embedding requests never forward the chat-only `a0_api_mode` control to providers.
- Prompt Markdown may retain fenced JSON examples for readability; final system-prompt rendering removes only their JSON fence markers before model calls and preserves non-JSON fences.
- Copy live core-plugin changes back into tracked source under `plugins/`.
- Develop new custom plugins under ignored `usr/plugins/`; tracked bundled plugins live under `plugins/`.
diff --git a/models.py b/models.py
index e8cb488af..8730832f4 100644
--- a/models.py
+++ b/models.py
@@ -805,10 +805,12 @@ class LiteLLMEmbeddingWrapper(Embeddings):
configure_litellm()
apply_rate_limiter_sync(self.a0_model_conf, " ".join(inputs))
+ call_kwargs = _merge_litellm_call_kwargs(self.kwargs)
+ call_kwargs.pop("a0_api_mode", None)
resp = embedding(
model=self.model_name,
input=inputs,
- **_merge_litellm_call_kwargs(self.kwargs),
+ **call_kwargs,
)
return [
item.get("embedding") if isinstance(item, dict) else item.embedding # type: ignore
diff --git a/plugins/_model_config/AGENTS.md b/plugins/_model_config/AGENTS.md
index 1c4dce812..8ea553924 100644
--- a/plugins/_model_config/AGENTS.md
+++ b/plugins/_model_config/AGENTS.md
@@ -49,6 +49,7 @@
- Preset editor reset actions must remove the user override through the preset API and refresh the open draft from bundled defaults.
- Preset rename, delete, and reset actions must repair scoped config and durable/live chat references; removed definitions fall back to `Default`.
- Migration must preserve existing definitions and distinct scoped model choices, back up replaced user files once, strip inline secrets, and remain idempotent.
+- Venice migration applies chat-only kwargs to chat slots and removes them from embedding slots without discarding embedding-specific kwargs.
- On every startup after migration, short-circuit when `usr/plugins/_model_config/presets.yaml` exists. Only a missing collection may fetch `agent0ai/a0-presets`; parse remote and plugin-local fallback YAML through the same validator, strip secrets before persistence, and persist `mode_presets_fallback.yaml` when download or validation fails.
- Model-name catalogs open below the input from either a field click or the embedded magnifier.
diff --git a/plugins/_model_config/extensions/python/startup_migration/_10_migrate_model_config.py b/plugins/_model_config/extensions/python/startup_migration/_10_migrate_model_config.py
index 7e2b2f5dc..1ad2d0406 100644
--- a/plugins/_model_config/extensions/python/startup_migration/_10_migrate_model_config.py
+++ b/plugins/_model_config/extensions/python/startup_migration/_10_migrate_model_config.py
@@ -374,22 +374,47 @@ class MigrateModelConfig(Extension):
def _repair_venice_config_slots(self, config: dict) -> bool:
changed = False
for section in self.CONFIG_SECTIONS:
- changed = self._repair_venice_slot(config.get(section)) or changed
+ changed = self._repair_venice_slot(
+ config.get(section),
+ embedding=section == "embedding_model",
+ ) or changed
return changed
def _repair_venice_preset(self, preset: dict) -> bool:
changed = False
for section in self.PRESET_SECTIONS:
- changed = self._repair_venice_slot(preset.get(section)) or changed
+ changed = self._repair_venice_slot(
+ preset.get(section),
+ embedding=section == "embedding",
+ ) or changed
if not any(section in preset for section in self.PRESET_SECTIONS):
changed = self._repair_venice_slot(preset) or changed
return changed
- def _repair_venice_slot(self, slot) -> bool:
+ def _repair_venice_slot(self, slot, *, embedding: bool = False) -> bool:
if not isinstance(slot, dict):
return False
provider = str(slot.get("provider") or "").strip().lower()
- if provider != "venice" or slot.get("kwargs") == self.VENICE_KWARGS:
+ if provider != "venice":
return False
- slot["kwargs"] = deepcopy(self.VENICE_KWARGS)
+
+ kwargs = slot.get("kwargs")
+ if embedding:
+ if "kwargs" not in slot:
+ return False
+ repaired = (
+ {
+ key: value
+ for key, value in kwargs.items()
+ if key not in self.VENICE_KWARGS
+ }
+ if isinstance(kwargs, dict)
+ else {}
+ )
+ else:
+ repaired = self.VENICE_KWARGS
+
+ if kwargs == repaired:
+ return False
+ slot["kwargs"] = deepcopy(repaired)
return True
diff --git a/tests/test_model_config_api_keys.py b/tests/test_model_config_api_keys.py
index f0441366a..6b64aa662 100644
--- a/tests/test_model_config_api_keys.py
+++ b/tests/test_model_config_api_keys.py
@@ -433,7 +433,10 @@ def test_model_config_migration_repairs_saved_venice_user_slots(monkeypatch, tmp
"embedding_model": {
"provider": "venice",
"name": "embed",
- "kwargs": {},
+ "kwargs": {
+ **expected,
+ "dimensions": 1024,
+ },
},
}
),
@@ -456,6 +459,14 @@ def test_model_config_migration_repairs_saved_venice_user_slots(monkeypatch, tmp
"name": "proxy",
"kwargs": {"keep": True},
},
+ "embedding": {
+ "provider": "venice",
+ "name": "embed",
+ "kwargs": {
+ **expected,
+ "dimensions": 512,
+ },
+ },
},
{
"name": "Legacy raw preset",
@@ -477,10 +488,11 @@ def test_model_config_migration_repairs_saved_venice_user_slots(monkeypatch, tmp
assert config == {"model_preset": "Default"}
assert presets[0]["name"] == "Default"
assert presets[0]["chat"]["kwargs"] == expected
- assert presets[0]["embedding"]["kwargs"] == expected
+ assert presets[0]["embedding"]["kwargs"] == {"dimensions": 1024}
assert presets[0]["utility"]["kwargs"] == {"a0_api_mode": "responses"}
assert presets[1]["chat"]["kwargs"] == expected
assert presets[1]["utility"]["kwargs"] == {"keep": True}
+ assert presets[1]["embedding"]["kwargs"] == {"dimensions": 512}
assert presets[2]["chat"]["kwargs"] == expected
assert (plugin_dir / "config.json.pre-unified-presets.bak").exists()
diff --git a/tests/test_multimodal_embeddings.py b/tests/test_multimodal_embeddings.py
index f9561cbb3..e2da04844 100644
--- a/tests/test_multimodal_embeddings.py
+++ b/tests/test_multimodal_embeddings.py
@@ -3,7 +3,7 @@ from types import SimpleNamespace
import models
-def test_embed_forwards_provider_ready_inputs(monkeypatch):
+def test_embed_forwards_provider_ready_inputs_without_chat_api_mode(monkeypatch):
calls: list[dict] = []
def fake_embedding(**kwargs):
@@ -17,9 +17,11 @@ def test_embed_forwards_provider_ready_inputs(monkeypatch):
monkeypatch.setattr(models, "embedding", fake_embedding)
monkeypatch.setattr(models, "configure_litellm", lambda: None)
- monkeypatch.setattr(models, "_merge_litellm_call_kwargs", lambda kwargs: kwargs)
+ monkeypatch.setattr(models, "_merge_litellm_call_kwargs", lambda kwargs: dict(kwargs))
monkeypatch.setattr(models, "apply_rate_limiter_sync", lambda *_args: None)
- wrapper = models.LiteLLMEmbeddingWrapper("gemini-embedding-2", "gemini")
+ wrapper = models.LiteLLMEmbeddingWrapper(
+ "gemini-embedding-2", "gemini", a0_api_mode="responses"
+ )
inputs = [
"plain text",
"data:image/png;base64,image",