In DeerFlow, we currently only support non-reasoning models. This means models like OpenAI's o1/o3 or DeepSeek's R1 are not supported yet, but we plan to add support for them in the future. Additionally, all Gemma-3 models are currently unsupported due to the lack of tool usage capabilities.
`doubao-1.5-pro-32k-250115`, `gpt-4o`, `qwen-max-latest`,`qwen3-235b-a22b`,`qwen3-coder`, `gemini-2.0-flash`, `deepseek-v3`, and theoretically any other non-reasoning chat models that implement the OpenAI API specification.
You can switch the model in use by modifying the `conf.yaml` file in the root directory of the project, using the configuration in the [litellm format](https://docs.litellm.ai/docs/providers/openai_compatible).
---
### How to use OpenAI-Compatible models?
DeerFlow supports integration with OpenAI-Compatible models, which are models that implement the OpenAI API specification. This includes various open-source and commercial models that provide API endpoints compatible with the OpenAI format. You can refer to [litellm OpenAI-Compatible](https://docs.litellm.ai/docs/providers/openai_compatible) for detailed documentation.
The following is a configuration example of `conf.yaml` for using OpenAI-Compatible models:
```yaml
# An example of Doubao models served by VolcEngine
DeerFlow supports native integration with Google AI Studio (formerly Google Generative AI) API. This provides direct access to Google's Gemini models with their full feature set and optimized performance.
To use Google AI Studio models, you need to:
1. Get your API key from [Google AI Studio](https://aistudio.google.com/app/apikey)
2. Set the `platform` field to `"google_aistudio"` in your configuration
3. Configure your model and API key
The following is a configuration example for using Google AI Studio models:
```yaml
# Google AI Studio native API (recommended for Google models)
BASIC_MODEL:
platform: "google_aistudio"
model: "gemini-2.5-flash" # or "gemini-1.5-pro" ,...
api_key: YOUR_GOOGLE_API_KEY # Get from https://aistudio.google.com/app/apikey
**Note:** The `platform: "google_aistudio"` field is required to distinguish from other providers that may offer Gemini models through OpenAI-compatible APIs.
### How to use models with self-signed SSL certificates?
If your LLM server uses self-signed SSL certificates, you can disable SSL certificate verification by adding the `verify_ssl: false` parameter to your model configuration:
```yaml
BASIC_MODEL:
base_url: "https://your-llm-server.com/api/v1"
model: "your-model-name"
api_key: YOUR_API_KEY
verify_ssl: false # Disable SSL certificate verification for self-signed certificates
```
> [!WARNING]
> Disabling SSL certificate verification reduces security and should only be used in development environments or when you trust the LLM server. In production environments, it's recommended to use properly signed SSL certificates.
DeerFlow supports the integration of OpenRouter models. You can refer to [litellm OpenRouter](https://docs.litellm.ai/docs/providers/openrouter). To use OpenRouter models, you need to:
1. Obtain the OPENROUTER_API_KEY from OpenRouter (https://openrouter.ai/) and set it in the environment variable.
2. Add the `openrouter/` prefix before the model name.
3. Configure the correct OpenRouter base URL.
The following is a configuration example for using OpenRouter models:
1. Configure OPENROUTER_API_KEY in the environment variable (such as the `.env` file)
```ini
OPENROUTER_API_KEY=""
```
2. Set the model name in `conf.yaml`
```yaml
BASIC_MODEL:
model: "openrouter/google/palm-2-chat-bison"
```
Note: The available models and their exact names may change over time. Please verify the currently available models and their correct identifiers in [OpenRouter's official documentation](https://openrouter.ai/docs).
DeerFlow supports the integration of Azure OpenAI chat models. You can refer to [AzureChatOpenAI](https://python.langchain.com/api_reference/openai/chat_models/langchain_openai.chat_models.azure.AzureChatOpenAI.html). Configuration example of `conf.yaml`:
### How to configure context length for different models
Different models have different context length limitations. DeerFlow provides a method to control the context length between different models. You can configure the context length between different models in the `conf.yaml` file. For example:
DeerFlow allows you to control which domains are included or excluded in Tavily search results through the configuration file. This helps improve search result quality and reduce hallucinations by focusing on trusted sources.
* Filter low-quality content: Filter out results with low relevance scores
* Clear base64 encoded images
* Length truncation: Truncate each search result according to the user-configured length
The filtering of low-quality content and length truncation depend on user configuration, providing two configurable parameters:
* min_score_threshold: Minimum relevance score threshold, search results below this threshold will be filtered. If not set, no filtering will be performed;
* max_content_length_per_page: Maximum length limit for each search result content, parts exceeding this length will be truncated. If not set, no truncation will be performed;
These two parameters can be configured in `conf.yaml` as shown below:
```yaml
SEARCH_ENGINE:
engine: tavily
include_images: true
min_score_threshold: 0.4
max_content_length_per_page: 5000
```
That's meaning that the search results will be filtered based on the minimum relevance score threshold and truncated to the maximum length limit for each search result content.
DeerFlow allows you to disable web search functionality, which is useful for environments without internet access or when you want to use only local RAG knowledge bases.
### Configuration
You can disable web search in your `conf.yaml` file:
> If you disable web search, make sure to configure local RAG resources; otherwise, the researcher will operate in pure LLM reasoning mode without external data sources.
### Behavior When Web Search is Disabled
- **Background investigation**: Skipped entirely (relies on web search)
- **Researcher node**: Will use only RAG retriever tools if configured
- **Pure reasoning mode**: If no RAG resources are available, the researcher will rely solely on LLM reasoning
4. Turn it **ON** to enable multi-turn clarification. Clarification is **disabled** by default. You need to manually enable it through any of the above methods. When clarification is enabled, you'll see **"Max Clarification Rounds"** field appear below the toggle
6. Set the maximum number of clarification rounds (default: 3, minimum: 1)
7. Click **Save** to apply changes
**When enabled**, the Coordinator will ask up to the specified number of clarifying questions for vague topics before starting research, improving report relevance and depth. The `max_clarification_rounds` parameter controls how many rounds of clarification are allowed.
**Note**: The `max_clarification_rounds` parameter only takes effect when `enable_clarification` is set to `true`. If clarification is disabled, this parameter is ignored.
