> ## Documentation Index
> Fetch the complete documentation index at: https://docs.knowledgestack.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Model Compatibility

> Knowledge Stack is designed to work with multiple LLM providers through a unified proxy layer. This page explains how the system handles differences between model providers to ensure reliable agent...

***

## Supported Model Architecture

The agent accesses language models through a **LiteLLM proxy**, which provides a unified OpenAI-compatible API across multiple LLM providers. This means you can switch between different models by changing a single configuration value.

```
Your chosen model (Claude, GPT-4, Qwen, etc.)
  -> LiteLLM Proxy (unified API)
    -> Agent Framework
      -> Knowledge Stack Agent
```

***

## How Provider Differences Are Handled

Different LLM providers have varying behaviors around features like "thinking" tokens, tool calling formats, and response structures. Knowledge Stack includes a compatibility layer that handles these differences automatically.

### Thinking Token Compatibility

Some models (like certain Qwen models via OpenRouter) produce "thinking" or "reasoning" tokens during processing. These tokens show the model's internal reasoning but can cause compatibility issues:

* **Problem:** Some providers generate thinking tokens but reject them when sent back in follow-up requests
* **Solution:** The system is configured to not echo thinking tokens back to the provider, preventing this class of errors entirely

### Malformed Tool Call Handling

Occasionally, a model may produce a malformed tool call (for example, a tool call with an empty name). Rather than crashing the entire agent run:

* Malformed tool calls are automatically stripped from the response
* The agent framework's built-in retry mechanism gives the model another chance to respond correctly
* This is bounded by configurable retry limits to prevent infinite loops

### HTTP Error Recovery

If the model provider returns an HTTP 400 error (bad request) due to incompatible message history:

* The system automatically strips problematic content from the conversation history
* The request is retried once with the cleaned history
* If the retry also fails, the error is surfaced gracefully

***

## Graceful Degradation

Instead of crashing on provider issues, the system degrades gracefully:

1. Problematic responses are sanitized and retried
2. Retries are bounded (default: 3 output retries, 50 model requests maximum)
3. If all retries are exhausted, the agent run ends with an error message rather than an unhandled crash
4. The user always sees a response in their thread -- even if it is just "Something went wrong. Please try again."

***

## Configuration

| Setting             | Default                        | Description                                             |
| ------------------- | ------------------------------ | ------------------------------------------------------- |
| Model identifier    | `openai/agent-general-purpose` | The LiteLLM model alias to use                          |
| LiteLLM proxy URL   | `http://localhost:4000`        | URL of the LiteLLM proxy server                         |
| Output retries      | 3                              | Maximum retries when the model produces unusable output |
| Model request limit | 50                             | Maximum total LLM requests per agent run                |

To switch models, update the model identifier in your configuration. The LiteLLM proxy handles routing to the correct provider.