> ## Documentation Index
> Fetch the complete documentation index at: https://praison.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Lean Install (litellm-free)
> Use PraisonAI without litellm โ capability detection still works via a conservative static heuristic
On lean installs (no `litellm`), PraisonAI still detects what each model can do โ structured outputs, function calling, prompt caching, and web search โ via a conservative built-in heuristic.
```mermaid theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
graph LR
Agent[๐ค Agent] --> Helper[๐ supports_* helper]
Helper --> Check{litellm present?}
Check -->|Yes| LiteLLM[๐ litellm catalogue]
Check -->|No| Heuristic[๐ Static heuristic]
LiteLLM --> Decision[โ
Capability decision]
Heuristic --> Decision
classDef agent fill:#8B0000,stroke:#7C90A0,color:#fff
classDef helper fill:#189AB4,stroke:#7C90A0,color:#fff
classDef branch fill:#6366F1,stroke:#7C90A0,color:#fff
classDef fallback fill:#F59E0B,stroke:#7C90A0,color:#fff
classDef result fill:#10B981,stroke:#7C90A0,color:#fff
class Agent agent
class Helper helper
class Check branch
class LiteLLM helper
class Heuristic fallback
class Decision result
```
You installed PraisonAI without extra dependencies, and structured outputs still work.
## Quick Start
```bash theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
pip install praisonaiagents
```
```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
from praisonaiagents import Agent
from pydantic import BaseModel
class Recipe(BaseModel):
name: str
steps: list[str]
agent = Agent(name="Chef", instructions="Return a simple recipe.", llm="gpt-4o")
result = agent.chat("Give me a recipe for pancakes", output_pydantic=Recipe)
print(result)
```
Capability gating stays usable โ `gpt-4o` still reports structured-output support without `litellm`.
```bash theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
pip install litellm
```
```python theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
from praisonaiagents import Agent
from pydantic import BaseModel
class Recipe(BaseModel):
name: str
steps: list[str]
agent = Agent(name="Chef", instructions="Return a simple recipe.", llm="gpt-4o")
result = agent.chat("Give me a recipe for pancakes", output_pydantic=Recipe)
print(result)
```
Same code, now backed by litellm's live catalogue โ recommended for production and for newly-released models.
***
## How It Works
Each `supports_*` helper checks for `litellm` first, then falls back to the static heuristic only when litellm is absent.
```mermaid theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
sequenceDiagram
participant Agent
participant Helper as supports_* helper
participant LiteLLM as litellm
Agent->>Helper: supports_structured_outputs("gpt-4o")
Helper->>LiteLLM: _get_litellm()
alt litellm installed
LiteLLM-->>Helper: litellm module
Helper->>LiteLLM: litellm.supports_response_schema(...)
LiteLLM-->>Helper: authoritative result
else litellm not installed
LiteLLM-->>Helper: None
Helper->>Helper: run static heuristic
end
Helper-->>Agent: True / False
```
`litellm` stays authoritative whenever it is installed โ even when a helper is missing or raises, the result is `False` rather than the heuristic, so an unsupported param never reaches a provider request. The static heuristic runs **only** when `litellm` is `None`.
| Situation | Result |
| --------------------------------------------- | -------------------------------- |
| Empty `""` model name | `False` |
| litellm installed | litellm's answer (authoritative) |
| litellm installed but helper missing / raises | `False` (stays authoritative) |
| litellm not installed | Static heuristic |
| Unknown / novel model, no litellm | `False` |
***
## What each model gets on a lean install
These are the exact patterns from the static heuristic โ names are lowercased and any `provider/` prefix is stripped before matching.
| Model family | Structured outputs | Function calling | Parallel tools | Prompt caching | Native web search |
| ----------------------------------- | :----------------: | :--------------: | :------------: | :------------: | :---------------: |
| GPT-4o / GPT-4.1 / GPT-5 | โ
| โ
| โ
| โ
| โยน |
| o1 / o3 / o4 | โ
| โ
| โ
| โ | โ |
| GPT-3.5 | โ | โ
| โ
| โ | โ |
| Claude 3 / Sonnet / Opus / Haiku | โ
| โ
| โ
| โ
| โยฒ |
| Gemini 2.x | โ
| โ
| โ
| โ | โ
|
| Mistral / Mixtral / Llama 3 | โ | โ
| โยณ | โ | โ |
| Perplexity / Sonar | โ | โ | โ | โ | โ
|
| Grok 3 | โ | โ
| โ | โ | โ
|
| DeepSeek | โ | โ | โ | โ
| โ |
| Embeddings / Whisper / TTS / DALL-E | โ | โโด | โ | โ | โ |
ยน Only `*-search-preview` variants (e.g. `gpt-4o-search-preview`) report native web search.
ยฒ Claude uses `web_fetch` instead of native web search, so it is intentionally excluded from the web-search heuristic โ this is expected, not a bug.
ยณ Mistral / Mixtral / Llama / Grok report **serial** function calling but **not** parallel โ the parallel heuristic is deliberately narrower.
โด Function calling is disabled by design for `embedding`, `whisper`, `tts`, and `dall-e` model names.
The web-search heuristic also matches any model name containing the literal substring `search`, plus `gemini-2` and `grok-3`. Anthropic Claude is excluded on purpose.
***
## Lean or full install?
Use this to decide whether the lean install is enough or whether you should add `litellm`.
```mermaid theme={"theme":{"light":"vitesse-light","dark":"vitesse-dark"}}
graph TB
Q{Need authoritative detection for
new or non-standard model IDs?}
Q -->|Yes| Full[๐ Install litellm]
Q -->|No| Lean[๐ชถ Lean install is fine]
classDef question fill:#6366F1,stroke:#7C90A0,color:#fff
classDef full fill:#189AB4,stroke:#7C90A0,color:#fff
classDef lean fill:#10B981,stroke:#7C90A0,color:#fff
class Q question
class Full full
class Lean lean
```
## When to install `litellm` anyway
Add `litellm` when a wrong `True` would break a real request.
* **Newly released models** with no matching pattern โ the heuristic returns `False` until you add litellm.
* **Non-standard model IDs** โ custom deployments, self-hosted proxies, or renamed models.
* **Production workloads** where a wrong `True` would send an unsupported `response_format` or `web_search_options` param and fail the call.
The helpers live at `praisonaiagents.llm.model_capabilities`, but you never call them directly โ the Agent uses them for you. Just install `litellm` when you need authoritative detection.
***
## Best Practices
The static heuristic is conservative, not exhaustive. For production traffic, `pip install litellm` so capability detection matches each provider's live catalogue.
When you target a single provider and want a small footprint, skip `litellm`. Structured outputs, function calling, and prompt caching still gate correctly for mainstream models.
When `litellm` is installed but a helper is missing or raises, the result is `False` โ not the heuristic. Don't wrap these helpers with your own fallback; that would send unsupported params to the provider.
A novel model with no matching pattern reports `False` across the board on a lean install. Install `litellm` before shipping support for a new provider so its capabilities are detected.
***
## Related
A different feature โ the `praisonaiagents.lite` subpackage with `LiteAgent`, not the litellm-free capability fallback.
Optional dependency groups for bots, gateway, and storage.
Fall back across models โ also uses capability info.
Higher-level model-selection layer, distinct from these `supports_*` helpers.