Veriquery

Multimodal Agentic RAG with Graph-Enhanced Knowledge Base
for Hardware Datasheet Intelligent Q&A and Verification System

---

Demo

🎯 Overview

Veriquery is an end-to-end intelligent system for electronic component datasheet analysis, built on Multimodal Agentic RAG, knowledge graph enhancement, and formal reasoning engines. It transforms static PDF datasheets into queryable knowledge, enabling engineers to retrieve parameters, verify electrical compatibility, compare devices, and locate reference circuits through natural language interaction.

The system addresses a critical gap in electronic design workflows: the manual, error-prone process of extracting and cross-referencing specifications from heterogeneous datasheets. VeriQuery automates this pipeline with a hybrid retrieval architecture, a four-layer Electrical Rule Check (ERC) engine, and a Z-number augmented multi-criteria decision framework.

Unlike general-purpose RAG systems, VeriQuery is purpose-built for the datasheet domain — every core module incorporates domain-specific design: the ERC engine encodes JEDEC logic-level standards and Arrhenius thermal degradation models; the parameter extractor uses section-anchored regex targeting "Electrical Characteristics" tables; the hybrid retriever adds a structured path (SQLite FTS5) specifically to preserve tabular data integrity; and the scoring engine normalizes test conditions via semiconductor physics (CCM) rather than generic min-max scaling.

✨ Key Features

Module	Capability	Technology
Intelligent Q&A	Natural language queries over datasheet content with citation tracing	Agentic RAG + LangGraph + Hybrid Retrieval
Pinout Analysis	Automatic pin definition extraction and SVG diagram rendering	Knowledge Graph + Regex + LLM
ERC Check	Four-layer progressive electrical compatibility verification	Interval Arithmetic + JEDEC Standards
Parameter Comparison	Multi-device scoring with reliability-aware ranking	Z-number + B-SPOTIS + MEREC
Circuit Retrieval	Multi-modal circuit diagram search from PDF pages	CLIP + VLM + MaxSim
Document Management	PDF upload, parsing, indexing, and lifecycle management	PyMuPDF + Camelot + pdfplumber

🏗️ System Architecture

Core Algorithms

⭐ Three-Stage Parameter Extraction

Cascaded pipeline with decreasing confidence:

1. Structured Table Query — Direct lookup from extracted PDF tables (confidence: ~0.93)
2. Section-Anchored Regex — Pattern matching within "Electrical Characteristics" sections (confidence: 0.73–0.85)
3. Few-Shot LLM Verification — Targeted LLM extraction for remaining parameters (confidence: ~0.80)

Each stage processes only parameters missed by prior stages (cascaded fallback), ensuring high-confidence results are preserved first.

🌟 Multimodal Visual Extraction (CLIP Pre-filter + ColPali)

Circuit diagram retrieval from datasheet pages using a two-stage filtering pipeline:

Stage	Model	Function
L1: CLIP Zero-Shot	ViT-B/32	Fast pre-filtering — classifies each page as circuit/non-circuit via text-image similarity, eliminating ~70% of irrelevant pages before expensive VLM inference
L2: Patch Embedding	Qwen3.5-2B VLM	Deep feature extraction — splits images into patches, encodes each patch into multi-vector embeddings for fine-grained content matching

Retrieval uses MaxSim (ColPali late interaction): query tokens are compared against all image patch embeddings via dot-product, and the maximum score per query token is summed:

$$MaxSim(q, d) = \sum_{i=1}^{n_q} \max_{j=1}^{n_d} \langle q_i, d_j \rangle$$

This architecture avoids the information loss of single-vector global embedding — each image retains full patch-level granularity for precise schematic matching.

✨ Hybrid Retrieval with RRF Fusion

Three heterogeneous retrieval paths execute concurrently, with results merged via Reciprocal Rank Fusion (Cormack et al., 2009):

$$score(d) = \sum_{s} w_s \times \frac{1}{k + rank_s(d) + 1}, \quad k=60$$

Path	Method	Weight	Strength
Dense	Sentence-Transformer + ChromaDB (HNSW)	0.50	Semantic similarity
Sparse	BM25 + jieba tokenization	0.35	Exact keyword match
Structured	SQLite FTS5 + table store	0.15	Preserves table structure

All three paths execute concurrently via asyncio.gather with return_exceptions=True. Total latency equals max(T1, T2, T3) instead of T1+T2+T3. Individual path failures are tolerated — remaining paths still return results.

💫 Four-Layer ERC Engine

Progressive detection architecture for electrical compatibility verification:

Layer	Detection	Method	Reference
L1	Static Stability	JEDEC logic level + noise margin	JESD8 series
L2	Signal Integrity	Transmission line reflection analysis	Bogatin, E.
L3	Topology Conflict	Interface protocol + port attribute matrix	IEEE 1801 UPF
L4	Environmental Degradation	Interval arithmetic + Arrhenius model	Moore (1966), JESD22-A108D

Layer 4 uses Interval(lo, hi) primitives for uncertainty propagation through arithmetic operations, modeling temperature drift where electrical parameters expand from crisp values to intervals.

🌠 Three-Layer Parameter Scoring (CCM + Z-A-FoM + B-SPOTIS)

Layer	Function	Method
CCM	Test condition normalization	Semiconductor physics linear equivalent conversion
Z-A-FoM	Reliability fusion	Z-number (Zadeh, 2011) + Kang conversion
B-SPOTIS	Robust decision	MEREC objective weighting + SPOTIS (Dezert et al., 2020)

⚡ Agentic RAG Workflow (LangGraph DAG)

Intent-driven orchestration layer built on LangGraph's StateGraph, routing user queries through domain-specific processing pipelines:

START → intent_router ─┬─ qa → text_retrieval → response_generation → END
                       └─ pinout → pinout_node → END

Design choices grounded in engineering constraints:

• Regex-based intent routing over LLM classification — zero latency, zero cost, fully interpretable; sufficient for the current 2-intent space (qa / pinout). ERC check and device comparison bypass the router entirely since their inputs are structured parameters, not free-form natural language.
• DAG topology guarantee — LangGraph's acyclic graph ensures every workflow terminates; no infinite loops or stuck states.
• Graceful degradation — individual node failures return fallback responses rather than crashing the pipeline.
• Compiled graph reuse — the workflow is compiled once at startup (read-only, thread-safe) and invoked concurrently across sessions.

🚧 Performance & Evaluation (In Progress)

The quantitative evaluation of Veriquery is currently underway. We are conducting comprehensive benchmark tests across the following dimensions, and the detailed experimental results will be published in our upcoming academic paper:

• Extraction Accuracy: Evaluating the F1-score of our cascaded parameter extraction pipeline against baseline LLMs (e.g., direct prompt-based extraction).
• Retrieval Robustness: Measuring the Recall@K improvements brought by the RRF hybrid retrieval strategy on heterogeneous datasheet PDFs.
• Reasoning Reliability: Validating the accuracy of the 4-layer ERC engine against standard JEDEC/IEEE edge-case scenarios.

Stay tuned for the full technical report and evaluation dataset.

📁 Project Structure

veriquery/
├── api/                        # FastAPI backend
│   ├── main.py                 # Application entry, lifespan, middleware
│   ├── dependencies.py         # Service container, dependency injection
│   ├── error_handlers.py       # Global exception handlers
│   └── routers/
│       ├── chat.py             # Intelligent Q&A endpoints
│       ├── circuit.py          # Circuit retrieval endpoints
│       ├── compare.py          # Device comparison endpoints
│       ├── documents.py        # Document management endpoints
│       ├── erc.py              # ERC check endpoints
│       └── pinout.py           # Pinout analysis endpoints
├── agents/                     # LangGraph agent workflow
│   ├── workflow_graph.py       # DAG topology and intent routing
│   ├── workflow_nodes.py       # Intent routing, retrieval, generation
│   ├── comparison_node.py      # Multi-device comparison orchestration
│   └── erc_node.py             # ERC check orchestration
├── core/                       # Shared infrastructure
│   ├── config.py               # Pydantic Settings, singleton configuration
│   ├── schema.py               # Unified data models (AgentState, PinInfo, etc.)
│   ├── llm_client.py           # HuggingFace LLM client with quantization
│   ├── svg_renderer.py         # Pinout SVG diagram renderer
│   ├── memory_manager.py       # GPU memory management
│   ├── model_manager.py        # Model lifecycle management
│   ├── cleanup_manager.py      # Orphan data cleanup
│   ├── sqlite_utils.py         # SQLite health check and repair
│   └── exceptions.py           # Custom exception hierarchy
├── ingestion/                  # Document processing pipeline
│   ├── document_processor.py   # PDF parsing, CLIP filtering, table orchestration
│   └── image_indexer.py        # CLIP + VLM + MaxSim visual indexing
├── extraction/                 # Parameter and table extraction
│   ├── parameter_extractor.py  # Three-stage cascaded parameter extraction
│   └── table_extractor.py      # Three-layer table extraction (Camelot/pdfplumber)
├── retrieval/                  # Hybrid retrieval subsystem
│   ├── hybrid_retriever.py     # RRF fusion orchestrator
│   ├── vector_store.py         # ChromaDB dense retrieval
│   ├── bm25_store.py           # BM25 sparse retrieval
│   ├── table_store.py          # SQLite FTS5 structured retrieval
│   └── embeddings.py           # Sentence-Transformer embedding service
├── reasoning/                  # Formal reasoning engines
│   ├── erc_engine.py           # Four-layer ERC with interval arithmetic
│   └── parameter_scorer.py     # CCM + Z-A-FoM + B-SPOTIS scoring
├── knowledge/                  # Domain knowledge base
│   ├── graph_db.py             # SQLite knowledge graph schema (chips→pins→parameters)
│   ├── graph_query.py          # Knowledge graph query engine (3-level fallback)
│   ├── chip_importer.py        # Chip data import pipeline
│   └── pinout_library.py       # Built-in common chip pinout database
├── ui/                         # Streamlit frontend
│   ├── app.py                  # Main page with navigation cards
│   ├── api_client.py           # Backend API client
│   ├── theme.py                # Academic-style CSS theme
│   ├── sidebar_nav.py          # Sidebar navigation and document selector
│   └── pages/
│       ├── 1_Documents.py      # Document management page
│       ├── 2_Chat.py           # Intelligent Q&A page
│       ├── 3_Pinout.py         # Pinout analysis page
│       ├── 4_ERC.py            # ERC check page
│       ├── 5_Compare.py        # Parameter comparison page
│       └── 6_Circuit.py        # Circuit retrieval page
├── docs/                       # Screenshots and documentation assets
├── data/                       # Runtime data (gitignored)
├── pyproject.toml              # Project metadata and build config
├── requirements.txt            # Python dependencies
├── .env.example                # Environment variable template
├── logging.yaml                # Logging config template (reference)
└── start.ps1                   # Windows startup script

🛠️ Tech Stack

Category	Technology	Purpose
Backend Framework	FastAPI + Uvicorn	Async REST API with OpenAPI docs
Frontend Framework	Streamlit	Multi-page interactive UI
Agent Workflow	LangGraph	DAG-based stateful workflow orchestration
LLM	Qwen3.5 (HuggingFace)	Local inference with 4-bit quantization, natively multimodal
VLM	Qwen3.5 (HuggingFace)	Natively multimodal model for diagram understanding
Embedding	BGE / Qwen-Embedding (HuggingFace)	Dense text vectorization (1024-dim)
Vector DB	ChromaDB	HNSW approximate nearest neighbor search
Sparse Retrieval	rank_bm25 + jieba	BM25 keyword matching with Chinese tokenization
Structured Retrieval	SQLite + FTS5	Full-text search over table data
PDF Processing	PyMuPDF + pdfplumber + Camelot	Text extraction, table extraction, image rendering
Vision	CLIP + PIL	Image classification and filtering
Knowledge Graph	SQLite	Chip-pin-parameter relational storage
Configuration	Pydantic Settings	Type-safe env-based configuration
Logging	Python logging + RotatingFileHandler	Structured logging with rotation

🚀 Getting Started

Prerequisites

• Python 3.10+
• CUDA-capable GPU (recommended, 4GB+ VRAM for minimum models)
• Git

Installation

git clone https://github.com/FinalSunFlower/Veriquery.git
cd veriquery

python -m venv .venv

# Windows:
.venv\Scripts\activate
# Linux/macOS:
source .venv/bin/activate

pip install -r requirements.txt

Model Setup

All models are automatically downloaded from HuggingFace on first run. No manual download required. Configure which models to use in .env:

cp .env.example .env

Model	Config Key	Default	VRAM	Purpose
LLM	LLM_MODEL	Qwen/Qwen3.5-0.8B	~1GB	Text generation (natively multimodal)
VLM	VLM_MODEL	Qwen/Qwen3.5-2B	~2GB	Diagram understanding (natively multimodal)
Embedding	EMBEDDING_MODEL	BAAI/bge-large-zh-v1.5	~1GB	Text vectorization
CLIP	CLIP_MODEL	openai/clip-vit-base-patch32	~0.5GB	Image classification

VRAM recommendations by GPU:

GPU VRAM	Recommended LLM	Recommended VLM
4GB	Qwen/Qwen3.5-0.8B	Qwen/Qwen3.5-0.8B
8GB	Qwen/Qwen3.5-2B	Qwen/Qwen3.5-2B
12GB+	Qwen/Qwen3.5-4B	Qwen/Qwen3.5-4B

Tip: If you have limited VRAM, set EMBEDDING_DEVICE=cpu and LLM_QUANTIZE=true in .env.

Launch

Windows (automated):

.\start.ps1

Manual start:

# Terminal 1: Start backend API
python -m api.main

# Terminal 2: Start frontend UI
streamlit run ui/app.py --server.port 8501

Access the application:

• Frontend: http://localhost:8501
• Backend API: http://localhost:8000
• API Documentation: http://localhost:8000/docs
• Health Check: http://localhost:8000/health

Quick Start Guide

1. Upload a datasheet — Navigate to the Documents page and upload a PDF datasheet (e.g., NE5532, LM358)
2. Ask questions — Go to Chat and type natural language queries like "NE5532 supply voltage range?"
3. View pinouts — Use the Pinout page to see automatically generated SVG pin diagrams
4. Run ERC — Check electrical compatibility between driver and receiver chips
5. Compare devices — Select multiple devices for parameter comparison with scoring
6. Find circuits — Search for application circuit diagrams from uploaded datasheets

🔌 API Reference

The backend exposes RESTful endpoints under /api/v1:

Endpoint	Method	Description
/api/v1/documents/	GET	List documents
/api/v1/documents/upload	POST	Upload document
/api/v1/chat/	POST	Intelligent Q&A with RAG
/api/v1/chat/stream	POST	Streaming Q&A response
/api/v1/pinout/	POST	Pin definition extraction
/api/v1/erc/check	POST	Four-layer ERC compatibility check
/api/v1/compare/devices-enhanced	POST	Multi-device parameter comparison
/api/v1/circuit/search	POST	Multi-modal circuit diagram search
/health	GET	System health check

Full interactive documentation is available at /docs (Swagger UI) and /redoc (ReDoc).

🔧 Configuration Reference

All configuration is managed through environment variables (.env file), with sensible defaults in core/config.py. Key configuration groups:

Group	Key Variables	Default
LLM	LLM_MODEL, LLM_DEVICE, LLM_QUANTIZE	Qwen/Qwen3.5-0.8B, cuda, true
VLM	VLM_MODEL, VLM_QUANTIZE	Qwen/Qwen3.5-2B, true
Embedding	EMBEDDING_MODEL, EMBEDDING_DIMENSION	BAAI/bge-large-zh-v1.5, 1024
Retrieval	VECTOR_WEIGHT, BM25_WEIGHT, STRUCTURED_WEIGHT	0.50, 0.35, 0.15
Chunking	CHUNK_SIZE, CHUNK_OVERLAP	800, 200
Storage	CHROMA_PERSIST_DIR, DATA_DIR	./data/chroma, ./data

See .env.example for the complete list of configurable parameters.

💡 Design Principles

• Citation Tracing — Every factual statement carries a source citation (file, page, text snippet) for verification
• Graceful Degradation — Each subsystem initializes independently; single component failure does not crash the system
• Lazy Loading — GPU models load on first use, avoiding startup memory pressure
• Async Concurrency — Retrieval paths execute concurrently via asyncio.gather; total latency = max(T1, T2, T3)
• Singleton Pattern — Thread-safe model instances via double-checked locking
• Configuration Validation — Pydantic validates all settings at startup with clear error messages

Citation

If you use VeriQuery in your research, please cite:

@misc{veriquery2026,
  author       = {VeriQuery Team},
  title        = {VeriQuery: Multimodal Agentic RAG with Graph-Enhanced Knowledge Base for Hardware Datasheet Intelligent Q\&A and Verification System},
  year         = {2026},
  publisher    = {GitHub},
  journal      = {GitHub repository},
  howpublished = {\url{https://github.com/FinalSunFlower/Veriquery}}
}

License

This project is licensed under the MIT License — see the LICENSE file for details.