DeepSearcher combines cutting-edge LLMs (OpenAI o3, Qwen3, DeepSeek, Grok 4, Claude 4 Sonnet, Llama 4, QwQ, etc.) and Vector Databases (Milvus, Zilliz Cloud etc.) to perform search, evaluation, and reasoning based on private data, providing highly accurate answer and comprehensive report. This project is suitable for enterprise knowledge management, intelligent Q&A systems, and information retrieval scenarios.
- **Private Data Search**: Maximizes the utilization of enterprise internal data while ensuring data security. When necessary, it can integrate online content for more accurate answers.
- **Vector Database Management**: Supports Milvus and other vector databases, allowing data partitioning for efficient retrieval.
- **Flexible Embedding Options**: Compatible with multiple embedding models for optimal selection.
- **Multiple LLM Support**: Supports DeepSeek, OpenAI, and other large models for intelligent Q&A and content generation.
- **Document Loader**: Supports local file loading, with web crawling capabilities under development.
Install DeepSearcher using one of the following methods:
#### Option 1: Using pip
Create and activate a virtual environment(Python 3.10 version is recommended).
```bash
python -m venv .venv
source .venv/bin/activate
```
Install DeepSearcher
```bash
pip install deepsearcher
```
For optional dependencies, e.g., ollama:
```bash
pip install "deepsearcher[ollama]"
```
#### Option 2: Install in Development Mode
We recommend using [uv](https://github.com/astral-sh/uv) for faster and more reliable installation. Follow the [offical installation instructions](https://docs.astral.sh/uv/getting-started/installation/) to install it.
Clone the repository and navigate to the project directory:
```shell
git clone https://github.com/zilliztech/deep-searcher.git && cd deep-searcher
```
Synchronize and install dependencies:
```shell
uv sync
source .venv/bin/activate
```
For more detailed development setup and optional dependency installation options, see [CONTRIBUTING.md](CONTRIBUTING.md#development-environment-setup-with-uv).
### Quick start demo
To run this quick start demo, please prepare your `OPENAI_API_KEY` in your environment variables. If you change the LLM in the configuration, make sure to prepare the corresponding API key.
```python
from deepsearcher.configuration import Configuration, init_config
from deepsearcher.online_query import query
config = Configuration()
# Customize your config here,
# more configuration see the Configuration Details section below.
<p>The "LLMName" can be one of the following: ["DeepSeek", "OpenAI", "XAI", "SiliconFlow", "Aliyun", "PPIO", "TogetherAI", "Gemini", "Ollama", "Novita"]</p>
<p> The "Arguments dict" is a dictionary that contains the necessary arguments for the LLM class.</p>
<details>
<summary>Example (OpenAI)</summary>
<p> Make sure you have prepared your OPENAI API KEY as an env variable <code>OPENAI_API_KEY</code>.</p>
<p> You need to install together before running, execute: <code>pip install together</code>. More details about TogetherAI: https://www.together.ai/ </p>
</details>
<details>
<summary>Example (XAI Grok)</summary>
<p> Make sure you have prepared your XAI API KEY as an env variable <code>XAI_API_KEY</code>.</p>
<p> You need to install gemini before running, execute: <code>pip install google-genai</code>. More details about Gemini: https://ai.google.dev/gemini-api/docs </p>
</details>
<details>
<summary>Example (DeepSeek from PPIO)</summary>
<p> Make sure you have prepared your PPIO API KEY as an env variable <code>PPIO_API_KEY</code>. You can create an API Key <ahref="https://ppinfra.com/settings/key-management?utm_source=github_deep-searcher">here</a>. </p>
<p> More details about PPIO: https://ppinfra.com/docs/get-started/quickstart.html?utm_source=github_deep-searcher </p>
</details>
<details>
<summary>Example (Ollama)</summary>
<p> Follow <ahref="https://github.com/jmorganca/ollama">these instructions</a> to set up and run a local Ollama instance:</p>
<p><ahref="https://ollama.ai/download">Download</a> and install Ollama onto the available supported platforms (including Windows Subsystem for Linux).</p>
<p> View a list of available models via the <ahref="https://ollama.ai/library">model library</a>.</p>
<p> Fetch available LLM models via <code>ollama pull <name-of-model></code></p>
<p> Example: <code>ollama pull qwen3</code></p>
<p> To chat directly with a model from the command line, use <code>ollama run <name-of-model></code>.</p>
<p> By default, Ollama has a REST API for running and managing models on <ahref="http://localhost:11434">http://localhost:11434</a>.</p>
<p> Make sure you have prepared your Volcengine API KEY as an env variable <code>VOLCENGINE_API_KEY</code>. You can create an API Key <ahref="https://console.volcengine.com/ark/region:ark+cn-beijing/apiKey">here</a>. </p>
<p> You need to install zhipuai before running, execute: <code>pip install zhipuai</code>. More details about GLM: https://bigmodel.cn/dev/welcome </p>
</details>
<details>
<summary>Example (Amazon Bedrock)</summary>
<p> Make sure you have prepared your Amazon Bedrock API KEY as an env variable <code>AWS_ACCESS_KEY_ID</code> and <code>AWS_SECRET_ACCESS_KEY</code>.</p>
<p> You need to install boto3 before running, execute: <code>pip install boto3</code>. More details about Amazon Bedrock: https://docs.aws.amazon.com/bedrock/ </p>
</details>
<details>
<summary>Example (IBM watsonx.ai)</summary>
<p> Make sure you have prepared your watsonx.ai credentials as env variables <code>WATSONX_APIKEY</code>, <code>WATSONX_URL</code>, and <code>WATSONX_PROJECT_ID</code>.</p>
<p> You need to install ibm-watsonx-ai before running, execute: <code>pip install ibm-watsonx-ai</code>. More details about IBM watsonx.ai: https://www.ibm.com/products/watsonx-ai/foundation-models </p>
<p>The "EmbeddingModelName" can be one of the following: ["MilvusEmbedding", "OpenAIEmbedding", "VoyageEmbedding", "SiliconflowEmbedding", "PPIOEmbedding", "NovitaEmbedding"]</p>
<p> The "Arguments dict" is a dictionary that contains the necessary arguments for the embedding model class.</p>
<details>
<summary>Example (OpenAI embedding)</summary>
<p> Make sure you have prepared your OpenAI API KEY as an env variable <code>OPENAI_API_KEY</code>.</p>
<p> Use the built-in embedding model in Pymilvus, you can set the model name as <code>"default"</code>, <code>"BAAI/bge-base-en-v1.5"</code>, <code>"BAAI/bge-large-en-v1.5"</code>, <code>"jina-embeddings-v3"</code>, etc. <br/>
See [milvus_embedding.py](deepsearcher/embedding/milvus_embedding.py) for more details. </p>
<p> For Jina's embedding model, you need<code>JINAAI_API_KEY</code>.</p>
<p> You need to install pymilvus model before running, execute: <code>pip install pymilvus.model</code>. More details about Pymilvus: https://milvus.io/docs/embeddings.md </p>
</details>
<details>
<summary>Example (VoyageAI embedding)</summary>
<p> Make sure you have prepared your VOYAGE API KEY as an env variable <code>VOYAGE_API_KEY</code>.</p>
<p> You need to install voyageai before running, execute: <code>pip install voyageai</code>. More details about VoyageAI: https://docs.voyageai.com/embeddings/ </p>
<p> You need to install boto3 before running, execute: <code>pip install boto3</code>. More details about Amazon Bedrock: https://docs.aws.amazon.com/bedrock/ </p>
</details>
<details>
<summary>Example (Novita AI embedding)</summary>
<p> Make sure you have prepared your Novita AI API KEY as an env variable <code>NOVITA_API_KEY</code>.</p>
<p> More details about Novita AI: https://novita.ai/docs/api-reference/model-apis-llm-create-embeddings?utm_source=github_deep-searcher&utm_medium=github_readme&utm_campaign=link </p>
<p> You need to install zhipuai before running, execute: <code>pip install zhipuai</code>. More details about GLM: https://bigmodel.cn/dev/welcome </p>
<p> You need to install gemini before running, execute: <code>pip install google-genai</code>. More details about Gemini: https://ai.google.dev/gemini-api/docs </p>
<p> You need to install ollama before running, execute: <code>pip install ollama</code>. More details about Ollama Python SDK: https://github.com/ollama/ollama-python </p>
</details>
<details>
<summary>Example (PPIO embedding)</summary>
<p> Make sure you have prepared your PPIO API KEY as an env variable <code>PPIO_API_KEY</code>.</p>
<p> You need to install fastembed before running, execute: <code>pip install fastembed</code>. More details about fastembed: https://github.com/qdrant/fastembed </p>
<p> Make sure you have prepared your WatsonX credentials as env variables <code>WATSONX_APIKEY</code>, <code>WATSONX_URL</code>, and <code>WATSONX_PROJECT_ID</code>.</p>
<p> You need to install ibm-watsonx-ai before running, execute: <code>pip install ibm-watsonx-ai</code>. More details about IBM watsonx.ai: https://www.ibm.com/products/watsonx-ai/foundation-models </p>
Setting the <code>uri</code> as a local file, e.g. <code>./milvus.db</code>, is the most convenient method, as it automatically utilizes <ahref="https://milvus.io/docs/milvus_lite.md"target="_blank">Milvus Lite</a> to store all data in this file.
</li>
</ul>
<ul>
<li>
If you have a large-scale dataset, you can set up a more performant Milvus server using
<ahref="https://milvus.io/docs/quickstart.md"target="_blank">Docker or Kubernetes</a>.
In this setup, use the server URI, e.g., <code>http://localhost:19530</code>, as your <code>uri</code>.
You can also use any other connection parameters supported by Milvus such as <code>host</code>, <code>user</code>, <code>password</code>, or <code>secure</code>.
</li>
</ul>
<ul>
<li>
If you want to use <ahref="https://zilliz.com/cloud"target="_blank">Zilliz Cloud</a>,
the fully managed cloud service for Milvus, adjust the <code>uri</code> and <code>token</code>
according to the <ahref="https://docs.zilliz.com/docs/on-zilliz-cloud-console#free-cluster-details"
target="_blank">Public Endpoint and API Key</a> in Zilliz Cloud.
<p> Currently supported file types: please refer to the Docling documentation: https://docling-project.github.io/docling/usage/supported_formats/#supported-output-formats </p>
<p> You need to install docling before running, execute: <code>pip install docling</code>. More details about Docling: https://docling-project.github.io/docling/ </p>
<p> You need to install crawl4ai before running, execute: <code>pip install crawl4ai</code>. More details about Crawl4AI: https://docs.crawl4ai.com/ </p>
</details>
<details>
<summary>Example (Jina Reader)</summary>
<p> Make sure you have prepared your Jina Reader API KEY as an env variable <code>JINA_API_TOKEN</code> or <code>JINAAI_API_KEY</code>.</p>
<p> Currently supported file types: please refer to the Docling documentation: https://docling-project.github.io/docling/usage/supported_formats/#supported-output-formats </p>
<p> You need to install docling before running, execute: <code>pip install docling</code>. More details about Docling: https://docling-project.github.io/docling/ </p>
Example loading from url (*Set `FIRECRAWL_API_KEY` in your environment variables, see [FireCrawl](https://docs.firecrawl.dev/introduction) for more details*):
For more help information about a specific subcommand, you can use `deepsearcher [subcommand] --help`.
```shell
deepsearcher load --help
deepsearcher query --help
```
### Deployment
#### Configure modules
You can configure all arguments by modifying [config.yaml](./config.yaml) to set up your system with default modules.
For example, set your `OPENAI_API_KEY` in the `llm` section of the YAML file.
#### Start service
The main script will run a FastAPI service with default address `localhost:8000`.
```shell
$ python main.py
```
#### Access via browser
You can open url http://localhost:8000/docs in browser to access the web service.
Click on the button "Try it out", it allows you to fill the parameters and directly interact with the API.
---
## ❓ Q&A
**Q1**: Why I failed to parse LLM output format / How to select the LLM?
**A1**: Small LLMs struggle to follow the prompt to generate a desired response, which usually cause the format parsing problem. A better practice is to use large reasoning models e.g. deepseek-r1 671b, OpenAI o-series, Claude 4 sonnet, etc. as your LLM.
---
**Q2**:
OSError: We couldn't connect to 'https://huggingface.co' to load this file, couldn't find it in the cached files and it looks like GPTCache/paraphrase-albert-small-v2 is not the path to a directory containing a file named config.json.
Checkout your internet connection or see how to run the library in offline mode at 'https://huggingface.co/docs/transformers/installation#offline-mode'.
**A2**: This is mainly due to abnormal access to huggingface, which may be a network or permission problem. You can try the following two methods:
1. If there is a network problem, set up a proxy, try adding the following environment variable.
```bash
export HF_ENDPOINT=https://hf-mirror.com
```
2. If there is a permission problem, set up a personal token, try adding the following environment variable.
```bash
export HUGGING_FACE_HUB_TOKEN=xxxx
```
---
**Q3**: DeepSearcher doesn't run in Jupyter notebook.
**A3**: Install `nest_asyncio` and then put this code block in front of your jupyter notebook.
