OopenJiuwen-botfeat: 对主流平台上的常用混合模型关闭思考模式

`openjiuwen_deepsearch` directory layout

This document reflects the current deepsearch/openjiuwen_deepsearch tree and what each major area does.

Overview

openjiuwen_deepsearch/
├── algorithm/                      # Core algorithms
│   ├── prompts/                    # Prompt templates
│   ├── query_understanding/        # Query understanding (router/outliner/planner/interpreter)
│   ├── report/                     # Report generation
│   ├── report_template/            # Template parse/generate
│   ├── research_collector/         # Gathering and evaluation
│   ├── source_trace/               # Provenance and validation
│   ├── source_tracer_infer/        # Provenance reasoning
│   └── user_feedback_processor/    # Post-report local edits from user feedback
├── framework/                      # Framework integration
│   └── openjiuwen/
│       ├── agent/                  # Workflow and nodes
│       ├── core/                   # WorkflowAgent and controller
│       ├── tools/                  # Search tool wrappers
│       └── llm/                    # LLM factory
├── config/                         # Configuration
├── common/                         # Shared exceptions and status codes
├── utils/                          # Utilities
└── llm/                            # Unified LLM wrapper

Details

`algorithm/` — core algorithms

Functions: Core algorithm implementations for each stage of the research workflow.

Main subdirectories:

prompts/ - Prompt templates (.md)
- synonym_rewrite_expand.md - Prompt for expansion
- synonym_rewrite_polish.md - Prompt for polishing
- synonym_rewrite_shorten.md - Prompt for shortening
- supplementary_search_task.md - Prompt for supplementary-search task generation
- supplementary_search_rewrite_selected_only.md - Prompt for supplementary search that rewrites only the selected span
- supplementary_search_rewrite_selected_and_related.md - Prompt for supplementary search that rewrites the entire related section
- new_task_assessment.md - Prompt for assessing sufficiency of historical evidence for new-task requests
- new_task_rewrite_section.md - Prompt for rewriting sections based on new-task edit strategy
query_understanding/ - Query understanding
- interpreter.py - Generate clarification questions
- outliner.py - Generate outlines
- planner.py - Generate section plans
- router.py - Decide whether to enter deep search
report/ - Report generation
- report.py - Main report generation logic
- report_utils.py - Report utility functions
- config.py - Report style and formatting
report_template/ - Template generation and parsing
- template_generator.py
- template_utils.py
research_collector/ - Information collection and evaluation
- collector_function.py
- doc_evaluation.py
- tool_log.py
source_trace/ - Provenance module
- source_tracer.py
- checker.py
- add_source.py
- citation_checker_research.py
- citation_verify_research.py
- content_analyzer.py
- source_matcher.py
- source_tracer_preprocessors.py
source_tracer_infer/ - Provenance reasoning module
- generate_html.py
- html_template.py
- infer.py
- infer_call_model.py
- infer_extract_info.py
- number_node.py
- supplement_graph.py
user_feedback_processor/ - User-feedback local editing module
- action_definitions.py - Mapping between frontend actions and unified internal actions
- common.py - Shared utilities (session/model context resolution, LLM invocation entry)
- history.py - Rewrite history and outline update management
- new_task_processor.py - New-task action processing logic
- report_edit_utils.py - Tools for stripping citation / inference markers and updating offsets
- section_locator.py - Locate the smallest Markdown heading block for a selection
- supplementary_search.py - Execution logic for supplementary search and local / whole-section rewriting
- synonym_rewrite.py - Execution logic for expansion, polishing, and shortening
- user_feedback_processor.py - Parse feedback, validate, execute, and send results

`framework/` — orchestration

Functions: Workflow and node orchestration based on openjiuwen.

Main subdirectories:

openjiuwen/agent/ - Workflows and nodes
- workflow.py - Agent and workflow entry
- main_graph_nodes.py - Main graph nodes (Start/Entry/Outline/Reporter/SourceTracer, etc.)
- editor_team_manager_node.py - Editor-team subgraph manager
- reasoning_writing_graph/ - Editor-team subgraph nodes and state
  - editor_team_nodes.py
  - dependency_reasoning_team_nodes.py
  - dependency_writing_team_nodes.py
  - section_context.py
- collector_graph/ - Information-collection subgraph
  - collector_execution_service.py - Reusable information-collection execution service
  - graph_builder.py
  - info_collector.py
  - collector_context.py
- agent_factory.py - Agent factory
- base_node.py - Base class for nodes
- search_context.py - Search context model
openjiuwen/core/workflow_agent/ — WorkflowAgent & controller
- config.py, workflow_controller.py, workflow_agent.py
openjiuwen/tools/ — search tools
- web_search.py, local_search.py, search_api/ (external_tool/, petal/, tavily/, serper/, xunfei/, local_search_api/, native_local_search_api/)
openjiuwen/llm/ — LLM factory
- llm_model_factory.py, llm_adapter.py

`config/`

config.py — LLMConfig, AgentConfig, ServiceConfig, etc.
method.py — execution mode enum
search_mode.py — search mode enum

`common/`

common_constants.py
exception.py
status_code.py

`utils/`

common_utils/ — llm_utils.py, security_utils.py, stream_utils.py, text_utils.py, url_utils.py
constants_utils/ — node_constants.py, session_contextvars.py, search_engine_constants.py
debug_utils/ — node_debug.py, outline_visualization.py, result_exporter.py
log_utils/ — logging helpers
validation_utils/ — field_validation.py, param_validation.py
rate_limiter_utils/ — qps_limiter.py

`llm/`

llm_wrapper.py — unified LLM calls
llm_request_adapter.py — LLM request parameter adaptation, including provider rules for the model thinking-mode switch

Module relationships

User request
    ↓
framework/openjiuwen/agent/workflow.py
    ├── validate & merge agent_config
    ├── init LLM & search tools
    └── Runner.run_agent_streaming(...)
            ↓
framework/openjiuwen/agent/main_graph_nodes.py
    ├── StartNode
    ├── EntryNode → algorithm/query_understanding/router.py
    ├── [GenerateQuestionsNode -> FeedbackHandlerNode] (optional HITL)
    ├── OutlineNode / DependencyOutlineNode → algorithm/query_understanding/outliner.py
    ├── OutlineInteractionNode / DependencyOutlineInteractionNode (optional)
    ├── EditorTeamNode / DependencyReasoningTeamNode / DependencyWritingTeamNode
    │   ├── ResearchPlanReasoningNode → algorithm/query_understanding/planner.py
    │   ├── InfoCollectorNode → collector_graph/
    │   └── SubReporterNode → algorithm/report/report.py
    ├── ReporterNode → algorithm/report/report.py
    ├── SourceTracerNode → algorithm/source_trace/
    ├── SourceTracerInferNode → algorithm/source_tracer_infer/
    └── UserFeedbackProcessorNode → algorithm/user_feedback_processor/

Where to look

Workflow → framework/openjiuwen/agent/
Algorithms → algorithm/
Config → config/config.py
Web search backends → framework/openjiuwen/tools/search_api/
Prompts → algorithm/prompts/
Context model → framework/openjiuwen/agent/search_context.py

Design principles

Layering: algorithm/ = logic; framework/ = orchestration.
Modularity: nodes stay decoupled from algorithm details.
Configuration: config/ is the single place for tunables.
Reuse: utils/ holds shared infrastructure.

openjiuwen_deepsearch directory layout

Overview

Details

algorithm/ — core algorithms

framework/ — orchestration

config/

common/

utils/

llm/