KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #developer-tools #ai-agent #claude-code #codex #typescript #ai-agents #deepseek-harness #mcp #cli #rust

Vertical-Industry-Intelligent-Research-Copilot

★ 59 Open GitHub ↗

No description available.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Vertical Industry Intelligent Research Copilot #

# Vertical Industry Intelligent Research Copilot #

**A multi-agent deep research system for vertical-domain intelligence synthesis, evidence-grounded analysis, and long-form report generation**

</div>

------

## Abstract ##

**Vertical Industry Intelligent Research Copilot** is a research-oriented, multi-agent AI system designed for **complex vertical-domain analysis** in scenarios such as **smart transportation, fintech, healthcare, and energy**.  
Unlike generic chat assistants, the system is built for **end-to-end research workflows**: it decomposes open-ended analytical questions into a coordinated process of **planning, retrieval, evidence tracing, data extraction, visualization, report writing, and adversarial review**.

The project targets a common limitation of traditional RAG pipelines: while they can retrieve documents effectively, they often struggle with **deep task decomposition**, **multi-source evidence synthesis**, **structured analytical reasoning**, and **quality control for long-form outputs**. To address this, this system introduces a **multi-agent research graph** with explicit role specialization, controlled execution, and iterative refinement.

------

## Highlights ##

- **Multi-agent research workflow** for complex analytical tasks  
- **Dual-mode retrieval** across web intelligence and local knowledge bases  
- **Hypothesis-driven planning** instead of simple retrieval-and-stitch generation  
- **Source tracing** to locate original references behind secondary claims  
- **Structured data extraction + chart generation** from unstructured materials  
- **Safe Python sandbox** with self-healing code execution  
- **Adversarial review loop** for factuality, coherence, and completeness  
- **Checkpoint & resume** for long-running research sessions

------

## Problem Statement ##

Vertical-domain research tasks are fundamentally different from simple question answering.  
A useful research assistant must do more than retrieve top-k passages. It must:

1. **understand the analytical objective**
2. **decompose the task into sub-problems**
3. **gather evidence from heterogeneous sources**
4. **extract structured signals from noisy text**
5. **generate charts and intermediate analytical artifacts**
6. **write coherent reports instead of stitched summaries**
7. **audit its own output for unsupported or low-quality claims**

Conventional one-shot RAG systems often fall short because they are prone to:

- shallow retrieval aggregation
- weak planning for long-form analysis
- poor source grounding for quantitative claims
- limited support for structured data workflows
- brittle output quality in multi-stage research settings

This project is an attempt to build a more **research-native AI system**.

------

## System Overview ##

The system orchestrates six specialized agents:

- **ChiefArchitect**  
  Defines the research plan, produces hypotheses, and structures the analytical agenda.
- **DeepScout**  
  Performs web search, recursive exploration, body extraction, source scoring, and source tracing.
- **DataAnalyst**  
  Extracts structured indicators from unstructured evidence and prepares chart-ready data.
- **CodeWizard**  
  Executes Python code in a restricted sandbox for analysis and visualization.
- **LeadWriter**  
  Produces long-form structured reports from collected evidence and intermediate artifacts.
- **CriticMaster**  
  Reviews drafts adversarially across factuality, source quality, consistency, and completeness.

Together, they form a full research pipeline:

> **Planning → Retrieval → Analysis → Writing → Review → Revision**

------

## Architecture ##

```text
User Query
   │
   ▼
ChiefArchitect
   │
   ├── Research agenda construction
   ├── Hypothesis generation
   └── Task decomposition
   │
   ▼
DeepScout
   │
   ├── Web search (Bocha)
   ├── Recursive exploration
   ├── Web body extraction (trafilatura)
   ├── Source scoring
   └── Source tracing
   │
   ▼
DataAnalyst
   │
   ├── Market size / growth / share extraction
   ├── Timeline & comparison signal extraction
   └── ECharts-ready config construction
   │
   ▼
CodeWizard
   │
   ├── Safe Python execution
   ├── Data cleaning / aggregation
   ├── PNG chart generation
   └── Self-healing retries
   │
   ▼
LeadWriter
   │
   ├── Report drafting
   ├── Section organization
   └── Evidence-grounded synthesis
   │
   ▼
CriticMaster
   │
   ├── Factuality checks
   ├── Source reliability checks
   ├── Logic / consistency checks
   ├── Missing evidence detection
   └── Revision routing
   │
   ├── major issue  ──► return to Retrieval / Revision
   └── pass         ──► Complete
```

------

## Core Design Choices ##

### 1. Hypothesis-Driven Research ###

Rather than simply retrieving documents for a keyword, the system first builds an **analytical agenda** and proposes **working hypotheses**.

Examples:

- “The fintech market will likely continue strong growth through 2026, but growth rates may vary by region.”
- “Embedded finance may become a stronger revenue driver in mature markets than basic digital payments.”

These hypotheses guide retrieval and evidence evaluation.  
This turns the workflow from **keyword accumulation** into **evidence-seeking research**.

------

### 2. Dual-Mode Retrieval ###

The system integrates two complementary retrieval channels:

#### Web Research Mode ####

Designed for up-to-date intelligence gathering.

- **Bocha Web Search API**
- **Recursive search** for following new evidence chains
- **Web content extraction** via `trafilatura`
- **Credibility scoring** for sources
- **Source tracing** for locating original references

#### Local Knowledge Base Mode ####

Designed for organization-specific or domain-specific corpora.

- **Milvus** vector database
- **text-embedding-v4** embeddings
- multi-format document support:
  - PDF
  - Word
  - Excel
  - PowerPoint
  - TXT / Markdown / HTML
- chunked indexing with context preservation

This design allows the system to combine **fresh external signals** with **internal domain knowledge**.

------

### 3. Structured Analysis Instead of Pure Text Generation ###

Research tasks frequently require converting narrative content into structured evidence.  
The system explicitly supports extraction of:

- market size
- growth rates / CAGR
- regional breakdowns
- segment shares
- company comparison signals
- policy-event timelines
- risk / opportunity factors

These signals can be turned into:

- **ECharts configurations** for frontend rendering
- **Python-generated PNG charts** through sandbox execution

Supported visual outputs include:

- line charts
- bar charts
- pie charts
- radar charts
- Sankey diagrams
- knowledge graphs

------

### 4. Safe Code Execution with Self-Healing ###

Only **CodeWizard** has execution privileges.

The sandbox is designed with:

- **module whitelisting**
- **forbidden pattern checks**
- **isolated global scope**
- **pre-execution syntax checking**
- **runtime error capture**
- **LLM-assisted repair and retry**

This enables lightweight analysis and plotting while maintaining a controlled execution boundary.

------

### 5. Adversarial Review as a First-Class Component ###

A key design principle is that long-form research generation should not terminate at first draft.  
**CriticMaster** acts as a red-team reviewer and evaluates outputs across:

- source reliability
- factual consistency
- logical coherence
- information completeness
- bias / one-sidedness
- hallucination risk

The workflow can then:

- return to retrieval for missing evidence
- return to revision for textual or structural problems
- finalize only when severe issues are resolved

This introduces an explicit **quality control loop** into the system.

------

### 6. Checkpointed Long-Running Research ###

Complex research tasks can be lengthy.  
The system therefore supports **checkpointing and resumption**, preserving:

- collected sources
- intermediate facts
- extracted data points
- report drafts
- workflow state

This makes the system more robust in multi-stage research sessions.

------

## Typical Use Cases ##

### Vertical Industry Research ###

- market sizing
- competitive landscape analysis
- value chain mapping
- technology trend tracking

### Competitive Intelligence ###

- company benchmarking
- product and capability comparison
- business model analysis
- strategic positioning

### Policy Impact Assessment ###

- regulatory change analysis
- policy-to-industry impact mapping
- regional policy comparison
- opportunity/risk identification

### Smart Transportation ###

- smart transportation market analysis
- V2X / vehicle-road-cloud integration studies
- smart highway research
- traffic policy and project opportunity tracking

### Long-Form Report Generation ###

- analyst-style market overviews
- company comparison reports
- policy briefs
- trend outlook reports

------

## Example Research Questions ##

- What is the projected global fintech market size in 2026, and how do forecast assumptions differ across sources?
- Compare Stripe, PayPal, and Adyen in terms of business model, positioning, and long-term competitive moat.
- Evaluate the impact of the EU AI Act on generative AI companies in Europe.
- Analyze the market potential, policy drivers, and technology routes of China’s smart transportation sector.
- Compare vehicle-road collaboration and single-vehicle intelligence as two development paths in smart mobility.
- Generate a structured deep research report on enterprise AI agents covering market, competitors, technology trends, policy environment, and risks.

------

## Prototype-Scale Output Profile ##

In typical research runs, the current prototype can produce:

- **20–40 min** average end-to-end research time depending on task complexity
- **20–40** effective sources aggregated per report
- **50–100** structured data points extracted
- **2–5** charts generated per task
- **3000–8000 words** of report output

> These numbers describe current prototype-level behavior and output scale, not a formal benchmark.

------

## Technical Stack ##

### Frontend ###

- React
- TypeScript
- Vite
- Ant Design
- ECharts

### Backend ###

- FastAPI
- Python
- SQLAlchemy
- PostgreSQL
- Redis

### Research Pipeline ###

- Multi-Agent Workflow
- RAG
- Milvus
- text-embedding-v4
- Bocha Web Search API
- trafilatura
- Python sandbox execution

### Infrastructure ###

- Docker / Docker Compose
- MinIO / etcd (for vector store dependencies)
- optional Elasticsearch integration

------

## Repository Structure ##

```text
industry_information_assistant/
├─ backend/
│  ├─ app/
│  │  ├─ router/
│  │  ├─ service/
│  │  │  ├─ deep_research_v2/
│  │  │  │  ├─ agents/
│  │  │  │  ├─ graph/
│  │  │  │  └─ state/
│  │  │  ├─ milvus_service.py
│  │  │  ├─ checkpoint_service.py
│  │  │  ├─ memory_service.py
│  │  │  └─ ...
│  │  ├─ model/
│  │  └─ app_main.py
│  └─ ...
├─ frontend/
│  ├─ src/
│  │  ├─ pages/
│  │  ├─ components/
│  │  ├─ services/
│  │  └─ ...
│  └─ ...
└─ ...
```

------

## Research Contributions ##

From a systems perspective, this project explores the following ideas:

1. **multi-agent decomposition for research tasks**  
   Long-form analysis is treated as a staged workflow rather than a single generation step.
2. **evidence-grounded synthesis over top-k retrieval**  
   The system seeks to organize evidence into analyst-style outputs instead of retrieval concatenation.
3. **bridging unstructured retrieval and structured analysis**  
   It supports extraction of quantitative signals and visual artifacts from textual materials.
4. **explicit reliability mechanisms for deep research**  
   Source tracing, adversarial review, sandboxing, and checkpointing are built into the pipeline.
5. **vertical-domain orientation**  
   The system is optimized for analytical domains where users care about research depth, source traceability, and decision support.

------

## Positioning ##

This project is best understood as:

> **a vertical-domain intelligent research copilot, not a general chatbot**

It is intended for tasks that require:

- heterogeneous evidence synthesis
- multi-step reasoning
- structured analytical output
- source-aware reporting
- long-form research generation

------

## Limitations ##

As a current prototype, the system still has important limitations:

- citation binding is not yet strictly sentence-level
- quantitative claims may still require stronger source-to-number alignment
- question scope control can sometimes drift toward full-industry reports
- evaluation remains prototype-level rather than benchmark-standard
- domain-specific ground-truth datasets are still under development

These limitations directly motivate future work.

------

## Future Directions ##

Planned next steps include:

- formal evaluation benchmarks for deep research systems
- stronger numeric consistency and source alignment checks
- sentence-level citation grounding
- improved question-type-aware planning
- better chart grounding and visual auditability
- richer domain-specific evaluation datasets
- stronger diversity control for source selection
- more rigorous report quality metrics

------

## Acknowledgement ##

This project was developed in a **challenge-led setting** with **external industry mentorship** and **academic supervision**.  
It is intended as a research-oriented systems prototype exploring how multi-agent pipelines can support **vertical-domain intelligence workflows**.

------

## License ##

This repository is currently intended for **research, portfolio, and demonstration purposes**.

If you plan to open-source it formally, consider adding a license such as:

- MIT
- Apache-2.0
- GPL-3.0

------

## Contact ##

If you are interested in:

- multi-agent systems
- deep research pipelines
- vertical-domain RAG
- intelligent report generation
- AI systems for industry intelligence

please feel free to connect through **GitHub Issues** or **Discussions**.

Vertical-Industry-Intelligent-Research-Copilot

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README