@samitouri / QOSami_Agents_Zero / commits / e2792e9c

refactor: Standardize model provider configuration

linuztx committed Jul 17, 2025 at 08:46 UTC e2792e9cfcf20d9c3e52eb7f079cf34a756c14cc
5 files changed +155 -90
conf/model_providers.yaml
+65 -30
@@ -1,55 +1,90 @@
1 # Supported model providers for Agent Zero
2 # ---------------------------------------
3 -# Each entry must contain:
4 -# id – identifier used in settings (lower-case, no spaces)
5 -# name – human readable name
6 -# Optional extra parameters are accepted (api_base, kwargs …)
7 -# Chat-capable and embedding-capable providers are listed separately as not every
8 -# provider exposes both kinds of models.
3 +#
4 +# Each provider type ("chat", "embedding") contains a mapping of provider IDs
5 +# to their configurations.
6 +#
7 +# The provider ID (e.g., "anthropic") is used:
8 +# - in the settings UI dropdowns.
9 +# - to construct the environment variable for the API key (e.g., ANTHROPIC_API_KEY).
10 +#
11 +# Each provider configuration requires:
12 +# name: Human-readable name for the UI.
13 +# litellm_provider: The corresponding provider name in LiteLLM.
14 +#
15 +# Optional fields:
16 +# kwargs: A dictionary of extra parameters to pass to LiteLLM.
17 +# This is useful for `api_base`, `extra_headers`, etc.
18
19 chat:
11 - - id: anthropic
20 + anthropic:
21 name: Anthropic
13 - - id: deepseek
22 + litellm_provider: anthropic
23 + deepseek:
24 name: DeepSeek
15 - - id: gemini
25 + litellm_provider: deepseek
26 + gemini:
27 name: Google
17 - - id: groq
28 + litellm_provider: gemini
29 + groq:
30 name: Groq
19 - - id: huggingface
31 + litellm_provider: groq
32 + huggingface:
33 name: HuggingFace
21 - - id: lm_studio
34 + litellm_provider: huggingface
35 + lm_studio:
36 name: LM Studio
23 - - id: mistral
37 + litellm_provider: lm_studio
38 + mistral:
39 name: Mistral AI
25 - - id: ollama
40 + litellm_provider: mistral
41 + ollama:
42 name: Ollama
27 - - id: openai
43 + litellm_provider: ollama
44 + openai:
45 name: OpenAI
29 - - id: azure
46 + litellm_provider: openai
47 + azure:
48 name: OpenAI Azure
31 - - id: openrouter
49 + litellm_provider: azure
50 + openrouter:
51 name: OpenRouter
33 - - id: sambanova
52 + litellm_provider: openrouter
53 + kwargs:
54 + extra_headers:
55 + "HTTP-Referer": "https://agent-zero.ai/"
56 + "X-Title": "Agent Zero"
57 + sambanova:
58 name: Sambanova
35 - - id: venice
59 + litellm_provider: sambanova
60 + venice:
61 name: Venice
37 - api_base: https://api.venice.ai/api/v1
38 - - id: other
62 + litellm_provider: openai
63 + kwargs:
64 + api_base: https://api.venice.ai/api/v1
65 + other:
66 name: Other OpenAI compatible
67 + litellm_provider: openai
68
69 embedding:
42 - - id: huggingface
70 + huggingface:
71 name: HuggingFace
44 - - id: mistral
45 - name: Mistral AI
46 - - id: lm_studio
72 + litellm_provider: huggingface
73 + lm_studio:
74 name: LM Studio
48 - - id: ollama
75 + litellm_provider: lm_studio
76 + mistral:
77 + name: Mistral AI
78 + litellm_provider: mistral
79 + ollama:
80 name: Ollama
50 - - id: openai
81 + litellm_provider: ollama
82 + openai:
83 name: OpenAI
52 - - id: azure
84 + litellm_provider: openai
85 + azure:
86 name: OpenAI Azure
54 - - id: other
55 - name: Other OpenAI compatible
\ No newline at end of file
87 + litellm_provider: azure
88 + other:
89 + name: Other OpenAI compatible
90 + litellm_provider: openai
\ No newline at end of file
models.py
+40 -44
@@ -351,6 +351,9 @@ class LocalSentenceTransformerWrapper(Embeddings):
351 """Local wrapper for sentence-transformers models to avoid HuggingFace API calls"""
352
353 def __init__(self, provider: str, model: str, **kwargs: Any):
354 + # Clean common user-input mistakes
355 + model = model.strip().strip('"').strip("'")
356 +
357 # Remove the "sentence-transformers/" prefix if present
358 if model.startswith("sentence-transformers/"):
359 model = model[len("sentence-transformers/") :]
@@ -449,20 +452,37 @@ def _adjust_call_args(provider_name: str, model_name: str, kwargs: dict):
452 if provider_name == "other":
453 provider_name = "openai"
454
452 - # Treat unknown providers that expose a custom OpenAI-compatible endpoint
453 - # (i.e. they pass an `api_base` URL) as generic OpenAI providers so that
454 - # LiteLLM can route the call correctly. This keeps dedicated providers
455 - # such as Azure and OpenRouter unchanged.
456 - if kwargs.get("api_base") and provider_name not in (
457 - "openai",
458 - "azure",
459 - "openrouter",
460 - ):
461 - provider_name = "openai"
462 -
455 return provider_name, model_name, kwargs
456
457
458 +def _merge_provider_defaults(
459 + provider_type: str, original_provider: str, kwargs: dict
460 +) -> tuple[str, dict]:
461 + provider_name = original_provider # default: unchanged
462 + cfg = get_provider_config(provider_type, original_provider)
463 + if cfg:
464 + provider_name = cfg.get("litellm_provider", original_provider).lower()
465 +
466 + # Extra arguments nested under `kwargs` for readability
467 + extra_kwargs = cfg.get("kwargs") if isinstance(cfg, dict) else None # type: ignore[arg-type]
468 + if isinstance(extra_kwargs, dict):
469 + for k, v in extra_kwargs.items():
470 + kwargs.setdefault(k, v)
471 +
472 + # Copy any additional top-level fields except metadata keys
473 + for k, v in cfg.items():
474 + if k not in ("id", "name", "value", "litellm_provider", "kwargs"):
475 + kwargs.setdefault(k, v)
476 +
477 + # Inject API key based on the *original* provider id if still missing
478 + if "api_key" not in kwargs:
479 + key = get_api_key(original_provider)
480 + if key and key not in ("None", "NA"):
481 + kwargs["api_key"] = key
482 +
483 + return provider_name, kwargs
484 +
485 +
486 def get_model(type: ModelType, provider: str, name: str, **kwargs: Any):
487 provider_name = provider.lower()
488 if type == ModelType.CHAT:
@@ -476,46 +496,22 @@ def get_model(type: ModelType, provider: str, name: str, **kwargs: Any):
496 def get_chat_model(
497 provider: str, name: str, **kwargs: Any
498 ) -> LiteLLMChatWrapper:
479 - provider_name = provider.lower()
480 -
481 - # Merge provider-specific defaults from configuration file
482 - cfg = get_provider_config("chat", provider_name)
483 - if cfg:
484 - extra = {k: v for k, v in cfg.items() if k not in ("id", "name", "value")}
485 - for k, v in extra.items():
486 - kwargs.setdefault(k, v)
487 -
488 - model = _get_litellm_chat(LiteLLMChatWrapper, name, provider_name, **kwargs)
489 - return model
499 + orig = provider.lower()
500 + provider_name, kwargs = _merge_provider_defaults("chat", orig, kwargs)
501 + return _get_litellm_chat(LiteLLMChatWrapper, name, provider_name, **kwargs)
502
503
504 def get_browser_model(
505 provider: str, name: str, **kwargs: Any
506 ) -> BrowserCompatibleChatWrapper:
495 - provider_name = provider.lower()
496 -
497 - cfg = get_provider_config("chat", provider_name)
498 - if cfg:
499 - extra = {k: v for k, v in cfg.items() if k not in ("id", "name", "value")}
500 - for k, v in extra.items():
501 - kwargs.setdefault(k, v)
502 -
503 - model = _get_litellm_chat(
504 - BrowserCompatibleChatWrapper, name, provider_name, **kwargs
505 - )
506 - return model
507 + orig = provider.lower()
508 + provider_name, kwargs = _merge_provider_defaults("chat", orig, kwargs)
509 + return _get_litellm_chat(BrowserCompatibleChatWrapper, name, provider_name, **kwargs)
510
511
512 def get_embedding_model(
513 provider: str, name: str, **kwargs: Any
514 ) -> LiteLLMEmbeddingWrapper | LocalSentenceTransformerWrapper:
512 - provider_name = provider.lower()
513 -
514 - cfg = get_provider_config("embedding", provider_name)
515 - if cfg:
516 - extra = {k: v for k, v in cfg.items() if k not in ("id", "name", "value")}
517 - for k, v in extra.items():
518 - kwargs.setdefault(k, v)
519 -
520 - model = _get_litellm_embedding(name, provider_name, **kwargs)
521 - return model
515 + orig = provider.lower()
516 + provider_name, kwargs = _merge_provider_defaults("embedding", orig, kwargs)
517 + return _get_litellm_embedding(name, provider_name, **kwargs)
preload.py
+2 -2
@@ -21,11 +21,11 @@ async def preload():
21
22 # preload embedding model
23 async def preload_embedding():
24 - if set["embed_model_provider"] == "HuggingFace":
24 + if set["embed_model_provider"].lower() == "huggingface":
25 try:
26 # Use the new LiteLLM-based model system
27 emb_mod = models.get_embedding_model(
28 - "HuggingFace", set["embed_model_name"]
28 + "huggingface", set["embed_model_name"]
29 )
30 emb_txt = await emb_mod.aembed_query("test")
31 return emb_txt
python/helpers/providers.py
+33 -8
@@ -24,20 +24,45 @@ class ProviderManager:
24 self._load_providers()
25
26 def _load_providers(self):
27 - """Loads provider configurations from the YAML file."""
27 + """Loads provider configurations from the YAML file and normalises them."""
28 try:
29 config_path = files.get_abs_path("conf/model_providers.yaml")
30 with open(config_path, "r", encoding="utf-8") as f:
31 - self._raw = yaml.safe_load(f) or {}
31 + raw_yaml = yaml.safe_load(f) or {}
32 except (FileNotFoundError, yaml.YAMLError):
33 - self._raw = {}
34 -
35 - # Build UI option lists (value / label) from raw data
33 + raw_yaml = {}
34 +
35 + # ------------------------------------------------------------
36 + # Normalise the YAML so that internally we always work with a
37 + # list-of-dicts [{id, name, ...}] for each provider type. This
38 + # keeps existing callers unchanged while allowing the new nested
39 + # mapping format in the YAML (id -> { ... }).
40 + # ------------------------------------------------------------
41 + normalised: Dict[str, List[Dict[str, str]]] = {}
42 +
43 + for p_type, providers in (raw_yaml or {}).items():
44 + items: List[Dict[str, str]] = []
45 +
46 + if isinstance(providers, dict):
47 + # New format: mapping of id -> config
48 + for pid, cfg in providers.items():
49 + entry = {"id": pid, **(cfg or {})}
50 + items.append(entry)
51 + elif isinstance(providers, list):
52 + # Legacy list format – use as-is
53 + items.extend(providers or [])
54 +
55 + normalised[p_type] = items
56 +
57 + # Save raw
58 + self._raw = normalised
59 +
60 + # Build UI-friendly option list (value / label)
61 self._options = {}
37 - for p_type, providers in (self._raw or {}).items():
62 + for p_type, providers in normalised.items():
63 opts: List[FieldOption] = []
39 - for p in providers or []:
40 - pid = (p.get("id") or p.get("value") or "").upper()
64 + for p in providers:
65 + pid = (p.get("id") or p.get("value") or "").lower()
66 name = p.get("name") or p.get("label") or pid
67 if pid:
68 opts.append({"value": pid, "label": name})
python/helpers/settings.py
+15 -6
@@ -500,8 +500,17 @@ def convert_out(settings: Settings) -> SettingsOutput:
500 # api keys model section
501 api_keys_fields: list[SettingsField] = []
502
503 - for provider in get_providers("chat"):
504 - api_keys_fields.append(_get_api_key_field(settings, provider["value"].lower(), provider["label"]))
503 + # Collect unique providers from both chat and embedding sections
504 + providers_seen: set[str] = set()
505 + for p_type in ("chat", "embedding"):
506 + for provider in get_providers(p_type):
507 + pid_lower = provider["value"].lower()
508 + if pid_lower in providers_seen:
509 + continue
510 + providers_seen.add(pid_lower)
511 + api_keys_fields.append(
512 + _get_api_key_field(settings, pid_lower, provider["label"])
513 + )
514
515 api_keys_section: SettingsSection = {
516 "id": "api_keys",
@@ -993,7 +1002,7 @@ def _write_sensitive_settings(settings: Settings):
1002 def get_default_settings() -> Settings:
1003 return Settings(
1004 version=_get_version(),
996 - chat_model_provider="OPENROUTER",
1005 + chat_model_provider="openrouter",
1006 chat_model_name="openai/gpt-4.1",
1007 chat_model_api_base="",
1008 chat_model_kwargs={"temperature": "0"},
@@ -1003,7 +1012,7 @@ def get_default_settings() -> Settings:
1012 chat_model_rl_requests=0,
1013 chat_model_rl_input=0,
1014 chat_model_rl_output=0,
1006 - util_model_provider="OPENROUTER",
1015 + util_model_provider="openrouter",
1016 util_model_name="openai/gpt-4.1-nano",
1017 util_model_api_base="",
1018 util_model_ctx_length=100000,
@@ -1012,13 +1021,13 @@ def get_default_settings() -> Settings:
1021 util_model_rl_requests=0,
1022 util_model_rl_input=0,
1023 util_model_rl_output=0,
1015 - embed_model_provider="HUGGINGFACE",
1024 + embed_model_provider="huggingface",
1025 embed_model_name="sentence-transformers/all-MiniLM-L6-v2",
1026 embed_model_api_base="",
1027 embed_model_kwargs={},
1028 embed_model_rl_requests=0,
1029 embed_model_rl_input=0,
1021 - browser_model_provider="OPENROUTER",
1030 + browser_model_provider="openrouter",
1031 browser_model_name="openai/gpt-4.1",
1032 browser_model_api_base="",
1033 browser_model_vision=True,