Model Provider Plugins

Map any model name to a provider id — built-in or plugin-registered via entry points

Model provider plugins map a bare model name (llama-3.1-70b, claude-sonnet-4-6, custom-model) to its canonical provider id (groq, anthropic, …) so runtime and config resolution work for every vendor — not just three hard-coded ones.

from praisonaiagents.llm import resolve_provider

resolve_provider("claude-sonnet-4-6")   # "anthropic"
resolve_provider("openai/gpt-4o")        # "openai"  (explicit prefix wins)
resolve_provider("llama-3.1-70b-groq")   # "groq"

Your agent picks a model by name; PraisonAI resolves the provider so providers.<name>.runtime_default applies automatically.

Quick Start

1. Resolve a built-in provider

   from praisonaiagents.llm import resolve_provider
   
   print(resolve_provider("gemini-1.5-pro"))   # "google"
   print(resolve_provider("deepseek-chat"))     # "deepseek"
   print(resolve_provider("grok-2"))            # "xai"

2. Use it from YAML runtime config

   framework: praisonai
   
   providers:
     groq:
       runtime_default: praisonai   # applies to any groq model
   
   agents:
     fast:
       role: Responder
       goal: Answer quickly
       instructions: "Answer in one sentence"
       llm: llama-3.1-70b-groq       # resolves to provider "groq"

How It Works

resolve_provider(model_name) returns the provider id in a fixed precedence.

Returns None when nothing matches.

Built-in providers

Publish a Plugin

Register a vendor out-of-tree — pip-installable, no PraisonAI code changes.

1. Declare the entry point

   # your-plugin/pyproject.toml
   [project.entry-points."praisonaiagents.model_providers"]
   bedrock = "my_plugin.matchers:is_bedrock_model"

   The entry-point name (bedrock) is the provider id resolve_provider(...) returns on a match — it must equal the key you use under providers: in YAML.

2. Write the matcher

   # my_plugin/matchers.py
   def is_bedrock_model(model_name_lowercased: str) -> bool:
       return (
           model_name_lowercased.startswith("bedrock/")
           or ".anthropic.claude-" in model_name_lowercased
       )

   The matcher receives the already-lowercased model name and returns bool.

3. Use it

   # after `pip install your-plugin`
   from praisonaiagents.llm import resolve_provider
   
   resolve_provider("bedrock/claude-3")   # "bedrock"

What Consumes This

praisonai_adapter._resolve_agent_runtime calls resolve_provider(agent_model) to pick providers.<name>.runtime_default. When no provider matches, a logger.debug line explains the miss and the provider-scoped default is not consulted:

runtime resolution: no provider registered for model '<model>'; provider-scoped 'runtime_default' not consulted

See Runtime Selection for the full resolution order.

Public API

# lazy re-export (preferred)
from praisonaiagents.llm import resolve_provider

# canonical module
from praisonaiagents.llm.model_providers import resolve_provider

Best Practices

Related