title	emoji	colorFrom	colorTo	sdk	sdk_version	app_file	pinned	license	models	short_description
VELA Research Agent	📊	blue	purple	gradio	5.12.0	app.py	false	mit	intrect/VELA	Korean Financial Research with 7B LLM

VELA Framework

Domain-Specialized LLM Research Agent for Korean Financial Markets

VELA is an open-source research agent framework that demonstrates how a single developer can build a domain-specialized LLM system competitive with $100M+ proprietary projects -- for under $235/month in compute costs.

Key Results

Metric	VELA (7B)	Qwen 2.5 7B Base	GPT-4o	Exaone 3.5 7.8B
Domain Knowledge (100pt)	87.5	72.0	81.0	74.5
Korean Fluency	Native	Mixed (CN leak)	Good	Native
Reasoning Trace	Structured	None	Free-form	None

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Training: SFT (58K samples) + DPO (26K pairs) on Korean financial domain
• Inference: 16 tok/s on Apple Silicon (MLX 4-bit), RunPod Serverless, or any vLLM server
• License: MIT

Architecture

User Query
    |
    v
[ResearchAgent] -- CoT Reasoning Loop (Think -> Search -> Analyze -> Conclude)
    |
    +-- [CoTReasoningEngine] -- TODO-based iterative reasoning with confidence gating
    +-- [ResearchSearchModule] -- Multi-source web search (Naver + DuckDuckGo)
    +-- [ContentExtractor] -- Web page & PDF content extraction
    +-- [AdversaryAgent] -- Cross-verification via Perplexity API (optional)
    |
    v
ResearchResult (structured JSON with trajectory, claim-evidence mapping)

Quick Start

# 1. Clone
git clone https://github.com/intrect/vela-framework.git
cd vela-framework

# 2. Install
pip install -e .

# 3. Configure
cp .env.example .env
# Edit .env with your API keys (RUNPOD_API_KEY, NAVER_CLIENT_ID, etc.)

# 4. Run
python inference.py --query "SK하이닉스 HBM 시장 전망" --backend mlx

Installation

Requirements

• Python 3.10+
• At least one LLM backend configured (RunPod, MLX, or vLLM)

From Source

pip install -e .

Dependencies

Core (auto-installed):

• pydantic>=2.0 -- Structured schemas
• requests -- HTTP client
• python-dotenv -- Environment configuration
• duckduckgo-search -- Web search fallback
• beautifulsoup4 -- Content extraction

Configuration

All configuration is via environment variables. Copy .env.example and fill in your keys:

# LLM Backends (configure at least one)
RUNPOD_API_KEY=your_key          # RunPod Serverless
RUNPOD_ENDPOINT_ID=your_endpoint
VELA_MLX_BASE_URL=http://localhost:8081/v1   # MLX server
VLLM_BASE_URL=http://localhost:8000/v1       # vLLM server

# Search APIs
NAVER_CLIENT_ID_0=your_id       # Naver Search API
NAVER_CLIENT_SECRET_0=your_secret

# Verification (optional)
PERPLEXITY_API_KEY=your_key     # Adversary Agent

Usage

Python API

from vela import ResearchAgent
from vela.schemas import ResearchOptions

# Initialize with your preferred backend
agent = ResearchAgent(llm_backend="mlx")

# Run research
result = agent.research(
    query="SK하이닉스 HBM 시장 전망",
    options=ResearchOptions(max_iterations=5),
)

# Access results
print(result.conclusion)
print(f"Confidence: {result.confidence:.0%}")
print(f"Sources: {len(result.sources)}")

# Save with full metadata (for training data generation)
from pathlib import Path
ResearchAgent.save_with_metadata(result, Path("output/result.json"))

CLI

# Basic research
python inference.py -q "삼성전자 반도체 전략" -b mlx

# With verification
python inference.py -q "네이버 AI 전략" --verify

# Save output
python inference.py -q "카카오 실적" -o result.json

# Verbose logging
python inference.py -q "현대차 전기차" -v

Web Interface

Try VELA directly in your browser via Gradio:

# Install with web dependencies
pip install -e ".[web]"

# Launch local demo
python app.py
# Opens at http://localhost:7860

Or try the hosted demo on HuggingFace Spaces.

LLM Backends

Backend	Use Case	Setup
runpod	Cloud GPU inference	Set RUNPOD_API_KEY + RUNPOD_ENDPOINT_ID
mlx	Apple Silicon local	Run MLX server, set VELA_MLX_BASE_URL
vllm	Any GPU server	Run vLLM, set VLLM_BASE_URL

How It Works

Chain-of-Thought Reasoning

VELA uses a TODO-based CoT protocol where each research iteration follows:

1. Think: Analyze current state and generate a TODO list
2. Search: Execute web searches (Naver + DuckDuckGo)
3. Analyze: Extract intermediate findings from collected sources
4. Conclude: Synthesize final report when confidence threshold is met

Reasoning Trace Format (Markdown)

**Step 1**:
**Thought**: SK하이닉스의 HBM 시장 점유율과 경쟁 구도 분석 필요
**Action**: search
**Query**: SK하이닉스 HBM3E 시장점유율 2025
**Confidence**: 35%

**Step 2**:
**Thought**: HBM 매출 비중과 영업이익률 데이터 확보 완료
**Action**: analyze
**Confidence**: 65%

**Step 3**:
**Thought**: 충분한 데이터 수집, 결론 도출 가능
**Action**: conclude
**Confidence**: 85%

Confidence Gating

The system uses a confidence gate at multiple levels:

• Per-step: Each reasoning step reports confidence (0-100%)
• Continuation: Research continues until confidence >= 80% or max iterations
• Synthesis: Final report includes overall confidence score

Adversary Verification (Optional)

When --verify is enabled, an independent AdversaryAgent cross-checks the research output using the Perplexity API, identifying:

• Factual inconsistencies
• Unsupported claims
• Missing counter-arguments

Training Your Own Model

VELA's training pipeline produced the domain-specialized model through:

1. SFT (Supervised Fine-Tuning): 58K samples of Korean financial analysis with structured reasoning traces
2. DPO (Direct Preference Optimization): 26K pairs targeting language purity (eliminating Chinese/English leaks from Qwen base) and reasoning quality

Model Weights

• HuggingFace: intrect/vela (GGUF Q4_K_M)
• Base: Qwen/Qwen2.5-7B-Instruct + LoRA (r=64, alpha=128)

Training Cost Breakdown

Component	Cost	Notes
RunPod RTX 4090	~$50/month	SFT + DPO training
Haiku API (data gen)	~$80	5 batches, 50K samples
Naver/Search APIs	~$30/month	Data collection
Perplexity API	~$20/month	Adversary verification
Total	~$235/month

Project Structure

vela-framework/
├── app.py               # Gradio web demo (HF Spaces)
├── inference.py          # CLI entry point
├── vela/                 # Core package
│   ├── agent.py          # ResearchAgent orchestrator
│   ├── reasoning.py      # CoT reasoning engine
│   ├── search.py         # Multi-source web search
│   ├── schemas.py        # Pydantic data models
│   ├── content_extractor.py
│   ├── adversary.py      # Verification agent
│   ├── config.py         # Centralized configuration
│   ├── prompts/          # System & research prompts
│   └── tools/            # LLM clients & utilities
│       ├── runpod_client.py
│       ├── mlx_client.py
│       ├── vllm_client.py
│       ├── ddg_search.py
│       ├── naver_search.py
│       ├── confidence_gate.py
│       └── fact_extractor.py
├── docs/
│   └── METHODOLOGY.md    # Detailed methodology
└── examples/
    └── simple_analysis.py

Methodology

See docs/METHODOLOGY.md for detailed documentation on:

• Reasoning Trace format specification
• CoT protocol design
• Training data generation pipeline
• DPO strategy for language purity
• Benchmark methodology

Known Limitations

This open-source demo uses public search APIs only. See Production Enhancements for commercial capabilities.

Category	Limitation	Impact	Production Note
Model Size	7B parameter model (Qwen2.5-7B base)	Complex multi-step reasoning may degrade compared to 70B+ models
Language	Korean financial domain only	English/multilingual queries produce lower quality output
Real-time Data	No direct market data feed (price, volume, orderbook)	Research relies on web search snippets, not live market data
Valuation	No financial database integration (e.g., FnGuide, Bloomberg)	Cannot provide real-time PER/PBR/EPS; relies on news-sourced figures	FnGuide integration available
Search Coverage	Naver News API + DuckDuckGo only	No access to paywalled sources (증권사 리포트, 유료 DB)	Securities firm reports in prod
Content Extraction	Top 3 sources per search step	Remaining sources provide title + snippet only (no full text)	Full-text extraction in prod
Inference Speed	~16 tok/s (MLX 4-bit) / ~5 tok/s (CPU BF16)	Full research cycle takes 30-120 seconds depending on iterations
Repetition	7B models may exhibit output repetition	Post-processing mitigates but does not fully eliminate
Confidence	Self-reported confidence (not calibrated)	Confidence scores reflect model's subjective estimate, not statistical accuracy
Temporal	Training data cutoff affects domain knowledge	Recent events after training may not be reflected in reasoning quality

Production Enhancements

In commercial deployments, VELA can integrate:

• FnGuide API: Real-time consensus, target prices, analyst ratings (50+ firms)
• Securities firm reports: Full-text extraction from major Korean brokerages
• Financial statements: 3+ years of balance sheet, cash flow, income statement
• Order flow data: Institutional/foreign investor net buying (real-time)

Contact hello@intrect.io for enterprise features.

What VELA is NOT

• Not a trading bot: VELA generates research reports, not trade signals or orders
• Not a financial advisor: Output is for informational/educational purposes only
• Not a real-time system: Research runs in batch mode (30-120s per query), not streaming
• Not a replacement for professional analysis: Designed to augment, not replace, human judgment

Contributing

Contributions are welcome. Please open an issue first to discuss what you would like to change.

License

MIT

Citation

@software{vela_framework_2026,
  title={VELA Framework: Domain-Specialized LLM Research Agent for Korean Financial Markets},
  author={intrect},
  year={2026},
  url={https://github.com/intrect/vela-framework}
}