migrate search, scrape tool into verl and add code to spin up retrieval server (#12)

Author: bachvudinh

examples/vllm_multiturn/config/search_multiturn_grpo.yaml

@@ -0,0 +1,23 @@
+hydra:
+  searchpath:
+    - file://verl/trainer/config
+
+defaults:
+  - ppo_trainer
+  - _self_
+
+data:
+  max_prompt_length: 1024
+  max_response_length: 1024
+  train_batch_size: 256
+  return_raw_chat: True
+  shuffle: False
+
+actor_rollout_ref:
+  hybrid_engine: True
+  rollout:
+    name: sglang
+    multi_turn:
+      enable: True
+      max_assistant_turns: 2
+      format: qwen

examples/vllm_multiturn/config/tool_config/geo3k_tool_config.yaml

@@ -0,0 +1,16 @@
+tools:
+  - class_name: "verl.tools.geo3k_tool.Geo3kTool"
+    config: 
+      type: native
+    tool_schema:
+      type: "function"
+      function:
+        name: "calc_geo3k_reward"
+        description: "A tool for calculating the reward of geo3k. (1.0 if parsed answer is correct, 0.0 if parsed answer is incorrect or not correctly parsed)"
+        parameters:
+          type: "object"
+          properties:
+            answer:
+              type: "string"
+              description: "The model's answer to the geo3k problem, must be a digits"
+          required: ["answer"]

examples/vllm_multiturn/config/tool_config/gsm8k_tool_config.yaml

@@ -0,0 +1,16 @@
+tools:
+  - class_name: "verl.tools.gsm8k_tool.Gsm8kTool"
+    config: 
+      type: native
+    tool_schema:
+      type: "function"
+      function:
+        name: "calc_gsm8k_reward"
+        description: "A tool for calculating the reward of gsm8k. (1.0 if parsed answer is correct, 0.0 if parsed answer is incorrect or not correctly parsed)"
+        parameters:
+          type: "object"
+          properties:
+            answer:
+              type: "string"
+              description: "The model's answer to the GSM8K math problem, must be a digits"
+          required: ["answer"]

examples/vllm_multiturn/config/tool_config/mcp_server.json

@@ -0,0 +1,8 @@
+{
+    "mcpServers": {
+        "Tavily Expert": {
+            "url": "your_tavily_expert_url",
+            "auth_token": "your_tavily_api_token"
+        }
+    }
+}

examples/vllm_multiturn/config/tool_config/mcp_tool_config.yaml

@@ -0,0 +1,11 @@
+tools:
+  - class_name: verl.tools.mcp_search_tool.MCPSearchTool
+    config:
+      rate_limit: 120
+      timeout: 120
+      type: mcp
+    mcp:
+      mcp_servers_config_path: ./mcp_server.json
+      # optional
+      tool_selected_list: 
+        - tavily_search_tool

examples/vllm_multiturn/config/tool_config/sandbox_fusion_tool_config.yaml

@@ -0,0 +1,24 @@
+tools:
+  - class_name: "verl.tools.sandbox_fusion_tools.SandboxFusionTool"
+    config: 
+      sandbox_fusion_url: "https://xxx.apigateway-cn-beijing.volceapi.com/run_code"
+      num_workers: 10
+      enable_global_rate_limit: true
+      rate_limit: 10
+      default_timeout: 30
+      default_language: "python"
+      memory_limit_mb: 1024
+      type: native
+
+    tool_schema:
+      type: "function"
+      function:
+        name: "code_interpreter"
+        description: "A tool for executing code."
+        parameters:
+          type: "object"
+          properties:
+            code:
+              type: "string"
+              description: "The code to execute."
+          required: ["code"]

examples/vllm_multiturn/config/tool_config/search_tool_config.yaml

@@ -0,0 +1,52 @@
+tools:
+  - class_name: verl.tools.janv2_tool.web_search_tool.WebSearchTool
+    config:
+      type: native
+      rag_server_url: "http://localhost:3030"
+      num_results: 10
+      topk_retrieval: 30
+      num_workers: 64
+      rate_limit: 100
+      timeout: 600
+      enable_global_rate_limit: true
+    tool_schema:
+      type: function
+      function:
+        name: web_search
+        description: |
+          Search for information using a query string. Returns a list of results 
+          with titles, URLs (document IDs), and preview snippets. Use the visit 
+          tool to retrieve full content of specific documents.
+        parameters:
+          type: object
+          properties:
+            query:
+              type: string
+              description: "The search query to find relevant documents"
+          required:
+            - query
+
+  - class_name: verl.tools.janv2_tool.scrape_tool.ScrapeTool
+    config:
+      type: native
+      rag_server_url: "http://localhost:3030"
+      num_workers: 50
+      rate_limit: 50
+      timeout: 600
+      enable_global_rate_limit: true
+    tool_schema:
+      type: function
+      function:
+        name: scrape
+        description: |
+          Visit a URL to retrieve the full content of a document. Use this after 
+          web_search to get complete document content. The URL should be in the 
+          format 'doc_<id>' as returned by web_search.
+        parameters:
+          type: object
+          properties:
+            url:
+              type: string
+              description: "The URL/document ID to visit (e.g., 'doc_123')"
+          required:
+            - url

examples/vllm_multiturn/local_dense_retriever/download.py

@@ -0,0 +1,44 @@
+# Copyright 2024 Bytedance Ltd. and/or its affiliates
+# Copyright 2023-2024 SGLang Team
+# Copyright 2025 Search-R1 Contributors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Adapted from https://github.com/PeterGriffinJin/Search-R1/blob/main/scripts/download.py
+
+
+import argparse
+
+from huggingface_hub import hf_hub_download
+
+parser = argparse.ArgumentParser(description="Download files from a Hugging Face dataset repository.")
+parser.add_argument("--repo_id", type=str, default="PeterJinGo/wiki-18-e5-index", help="Hugging Face repository ID")
+parser.add_argument("--save_path", type=str, required=True, help="Local directory to save files")
+
+args = parser.parse_args()
+
+repo_id = "PeterJinGo/wiki-18-e5-index"
+for file in ["part_aa", "part_ab"]:
+    hf_hub_download(
+        repo_id=repo_id,
+        filename=file,  # e.g., "e5_Flat.index"
+        repo_type="dataset",
+        local_dir=args.save_path,
+    )
+
+repo_id = "PeterJinGo/wiki-18-corpus"
+hf_hub_download(
+    repo_id=repo_id,
+    filename="wiki-18.jsonl.gz",
+    repo_type="dataset",
+    local_dir=args.save_path,
+)