DJ_service#

To further enhance the user experience of Data-Juicer, we have introduced a new service features based on API (Service API) and a MCP server, This enables users to integrate and utilize Data-Juicer’s powerful operator pool in a more convenient way. With these service features, users can quickly build data processing pipelines without needing to delve into the underlying implementation details of the framework, and seamlessly integrate with existing systems. Additionally, users can achieve environment isolation between different projects through this service. This document will provide a detailed explanation of how to launch and use this service feature, helping you get started quickly and fully leverage the potential of Data-Juicer.

API Service#

Start the Service#

Run the following code:

uvicorn service:app

API Calls#

The API supports calling all functions and classes in Data-Juicer’s __init__.py (including calling specific functions of a class). Functions are called via GET, while classes are called via POST.

Protocol#

URL Path#

For GET requests to call functions, the URL path corresponds to the function reference path in the Data-Juicer library. For example, from data_juicer.config import init_configs maps to the path data_juicer/config/init_configs. For POST requests to call a specific function of a class, the URL path is constructed by appending the function name to the class path in the Data-Juicer library. For instance, calling the compute_stats_batched function of the TextLengthFilter operator corresponds to the path data_juicer/ops/filter/TextLengthFilter/compute_stats_batched.

Parameters#

When making GET and POST calls, parameters are automatically converted into lists. Additionally, query parameters do not support dictionary transmission. Therefore, if the value of a parameter is a list or dictionary, we uniformly transmit it using json.dumps and prepend a special symbol <json_dumps> to distinguish it from regular strings.

Special Cases#
  1. For the cfg parameter, we default to transmitting it using json.dumps, without needing to prepend the special symbol <json_dumps>.

  2. For the dataset parameter, users can pass the path of the dataset on the server, and the server will load the dataset.

  3. Users can set the skip_return parameter. When set to True, the result of the function call will not be returned, avoiding errors caused by network transmission issues.

Function Calls#

GET requests are used for function calls, with the URL path corresponding to the function reference path in the Data-Juicer library. Query parameters are used to pass the function arguments.

For example, the following Python code can be used to call Data-Juicer’s init_configs function to retrieve all parameters of Data-Juicer:

import requests
import json

json_prefix = '<json_dumps>'
url = 'http://localhost:8000/data_juicer/config/init_configs'
params = {"args": json_prefix + json.dumps(['--config', './demos/process_simple/process.yaml'])}
response = requests.get(url, params=params)
print(json.loads(response.text))

The corresponding curl command is as follows:

curl -G "http://localhost:8000/data_juicer/config/init_configs" \
     --data-urlencode "args=--config" \
     --data-urlencode "args=./demos/process_simple/process.yaml"

Class Function Calls#

POST requests are used for class function calls, with the URL path constructed by appending the function name to the class path in the Data-Juicer library. Query parameters are used to pass the function arguments, while JSON fields are used to pass the arguments required for the class constructor.

For example, the following Python code can be used to call Data-Juicer’s TextLengthFilter operator:

import requests
import json

json_prefix = '<json_dumps>'
url = 'http://localhost:8000/data_juicer/ops/filter/TextLengthFilter/compute_stats_batched'
params = {'samples': json_prefix + json.dumps({'text': ['12345', '123'], '__dj__stats__': [{}, {}]})}
init_json = {'min_len': 4, 'max_len': 10}
response = requests.post(url, params=params, json=init_json)
print(json.loads(response.text))

The corresponding curl command is as follows:

curl -X POST \
  "http://localhost:8000/data_juicer/ops/filter/TextLengthFilter/compute_stats_batched?samples=%3Cjson_dumps%3E%7B%22text%22%3A%20%5B%2212345%22%2C%20%22123%22%5D%2C%20%22__dj__stats__%22%3A%20%5B%7B%7D%2C%20%7B%7D%5D%7D" \
  -H "Content-Type: application/json" \
  -d '{"min_len": 4, "max_len": 10}'

Note: If you need to call the run function of the Executor or Analyzer classes for data processing and data analysis, you must first call the init_configs or get_init_configs function to obtain the complete Data-Juicer parameters to construct these two classes. For more details, refer to the demonstration below.

Demonstration#

We have integrated AgentScope to enable users to invoke Data-Juicer operators for data cleaning through natural language. The operators are invoked via an API service. For the specific code, please refer to here.

MCP Server#

Overview#

The Data-Juicer MCP server provides data processing operators to assist in tasks such as data cleaning, filtering, deduplication, and more. To accommodate different use cases, we offer two server options:

  • Recipe-Flow: Allows filtering operators by type and tags, and supports combining multiple operators into a data recipe for execution.

  • Granular-Operators: Provides each operator as an independent tool, allowing you to flexibly specify a list of operators to use via environment variables, thus building a customized data processing pipeline.

Please note that the Data-Juicer MCP server features and available tools may change and expand as we continue to develop and improve the server.

The server supports two deployment methods: stdio and streamable-http. The stdio method does not support multiprocessing. If you require multiprocessing or multithreading capabilities, you must use the streamable-http deployment method. Configuration details for each method are provided below.

Recipe-Flow#

The Recipe-Flow mode provides the following MCP tools:

1. search_ops#

  • Search for available data processing operators with multiple search modes

  • Input:

    • query (str, optional): Search query string. Required for “keyword” and “bm25” modes

    • op_type (str, optional): The type of data processing operator to filter by (aggregator / deduplicator / filter / grouper / mapper / selector / pipeline)

    • tags (List[str], optional): A list of tags to filter operators (Modality: text / image / audio / video / multimodal; Resource: cpu / gpu; Model: api / vllm / hf)

    • match_all (bool): Whether all specified tags must match. Default is True

    • search_mode (str): Search strategy — “tags” (filter by type and tags, default), “regex” (regex pattern matching), or “bm25” (BM25 text relevance ranking)

    • top_k (int): Maximum number of results for “bm25” mode. Default is 10

  • Returns: A dictionary containing details about the matched operators

2. get_global_config_schema#

  • Dynamically retrieves the full schema of all available global configuration options (parameter name, type, default value, description)

  • No input parameters required

  • Returns: A config schema dictionary, useful for discovering what options can be passed to run_data_recipe and analyze_dataset via the extra_config parameter

3. get_dataset_load_strategies#

  • Dynamically retrieves all available dataset loading strategies and their configuration rules

  • No input parameters required

  • Returns: Information about each strategy including executor_type, data_type, data_source, required/optional fields, etc. Useful for understanding how to configure the dataset parameter in run_data_recipe and analyze_dataset for different data sources (local files, HuggingFace, S3, etc.)

4. run_data_recipe#

  • Executes a data processing recipe (equivalent to dj-process)

  • Input:

    • process (List[Dict]): A list of processing steps to execute, where each dictionary contains an operator name and its parameter dictionary

    • dataset_path (str, optional): The path to the dataset to be processed (simple mode for local files)

    • dataset (Dict, optional): Advanced dataset configuration supporting multiple data sources. Format: {"configs": [{"type": "local", "path": "..."}], "max_sample_num": 10000}. Use get_dataset_load_strategies to discover available options

    • export_path (str, optional): The path to export the processed dataset. Default is None, exporting to ‘./outputs’

    • np (int): Number of processes to use. Default is 1

    • extra_config (Dict, optional): Additional global configuration options. Use get_global_config_schema to discover all available options. Example: {"open_tracer": true, "trace_num": 20, "op_fusion": true}

  • Returns: A string representing the execution result

5. analyze_dataset#

  • Analyzes dataset quality distribution (equivalent to dj-analyze)

  • Computes statistics for specified filter/tagging operators, performs overall analysis, column-wise analysis, and correlation analysis, generating stats tables and distribution figures

  • Input:

    • process (List[Dict]): A list of filter/tagging operators to compute stats for

    • dataset_path (str, optional): The path to the dataset to be analyzed

    • dataset (Dict, optional): Advanced dataset configuration (same as run_data_recipe)

    • export_path (str, optional): The path to export the analysis results

    • np (int): Number of processes to use. Default is 1

    • percentiles (List[float], optional): Percentiles for distribution analysis. Default is [0.25, 0.5, 0.75]

    • extra_config (Dict, optional): Additional global configuration options (same as run_data_recipe)

  • Returns: A string indicating where the analysis results are saved

Typical Workflow#

  1. Search operators: Call search_ops to find suitable operators (supports tag filtering, keyword search, or BM25 natural language search)

  2. Discover configuration: Optionally call get_global_config_schema and get_dataset_load_strategies to discover available configuration options and data sources

  3. Analyze dataset: Call analyze_dataset to analyze dataset quality distribution and determine appropriate filter thresholds

  4. Execute processing: Based on the analysis results, call run_data_recipe to execute the actual data processing

Granular-Operators#

By default, this MCP server returns all Data-Juicer operator tools, each running independently.

To control the operator tools returned by the MCP server, specify the environment variable DJ_OPS_LIST_PATH:

  1. Create a .txt file.

  2. Add operator names to the file. For example:

text_length_filter
flagged_words_filter
image_nsfw_filter
text_pair_similarity_filter
  1. Set the path to the operator list as the environment variable DJ_OPS_LIST_PATH.

Configuration#

The following configuration examples demonstrate how to set up the two MCP server types using the stdio and SSE methods. These examples are for illustrative purposes only and should be adapted to the specific MCP client’s configuration format.

stdio#

Suitable for quick local testing and simple scenarios. Add the following to the MCP client’s configuration file (e.g., claude_desktop_config.json or similar):

Using uvx#

Run the latest version of Data-Juicer MCP directly from the repository without manual local installation.

  • Recipe-Flow mode:

    {
      "mcpServers": {
        "DJ_recipe_flow": {
          "command": "uvx",
          "args": [
            "--from",
            "git+https://github.com/datajuicer/data-juicer",
            "dj-mcp",
            "recipe-flow"
          ]
        }
      }
    }
    
  • Granular-Operators mode:

    {
      "mcpServers": {
        "DJ_granular_ops": {
          "command": "uvx",
          "args": [
            "--from",
            "git+https://github.com/datajuicer/data-juicer",
            "dj-mcp",
            "granular-ops",
            "--transport",
            "stdio"
          ],
          "env": {
            "DJ_OPS_LIST_PATH": "/path/to/ops_list.txt"
          }
        }
      }
    }
    

    Note: If DJ_OPS_LIST_PATH is not set, all operators are returned by default.

Local Installation#
  1. Clone the Data-Juicer repository locally:

    git clone https://github.com/datajuicer/data-juicer.git
    
  2. Run Data-Juicer MCP using uv:

  • Recipe-Flow mode:

    {
      "mcpServers": {
        "DJ_recipe_flow": {
          "transport": "stdio",
          "command": "uv",
          "args": [
            "run",
            "--directory",
            "/abs/path/to/data-juicer",
            "dj-mcp",
            "recipe-flow"
          ]
        }
      }
    }
    
  • Granular-Operators mode:

    {
      "mcpServers": {
        "DJ_granular_ops": {
          "transport": "stdio",
          "command": "uv",
          "args": [
            "run",
            "--directory",
            "/abs/path/to/data-juicer",
            "dj-mcp",
            "granular-ops"
          ],
          "env": {
            "DJ_OPS_LIST_PATH": "/path/to/ops_list.txt"
          }
        }
      }
    }
    

streamable-http#

To use streamable-http deployment, first start the MCP server separately.

  1. Run the MCP server: Execute the MCP server script and specify the port number:

    • Using uvx:

      uvx --from git+https://github.com/datajuicer/data-juicer dj-mcp <MODE: recipe-flow/granular-ops> --transport streamable-http --port 8080
      
    • Local execution:

      uv run dj-mcp <MODE: recipe-flow/granular-ops> --transport streamable-http --port 8080
      
  2. Configure your MCP client: Add the following to the MCP client’s configuration file:

    {
      "mcpServers": {
        "DJ_MCP": {
          "url": "http://127.0.0.1:8080/mcp"
        }
      }
    }
    

Notes:

  • URL: The url should point to the mcp endpoint of the running server (typically http://127.0.0.1:<port>/mcp). Adjust the port number if a different value was used when starting the server.

  • Separate server process: The streamable-http server must be running before the MCP client attempts to connect.

  • Firewall: Ensure the firewall allows connections to the specified port.