Pokemon-Chat

Overview#

Pokemon-Chat (可萌) is a domain-specific intelligent Q&A system for Pokémon enthusiasts, featuring LangGraph multi-agent orchestration and GraphRAG technologies. It addresses the limitations of general-purpose LLMs in vertical domains, including knowledge gaps and hallucination issues. The project provides a complete Docker Compose deployment solution and serves as a transferable template for domain-specific knowledge systems.

Core Features#

Intelligent Q&A and Retrieval#

• Knowledge Graph Q&A: Built on Neo4j, supporting structured relationship queries including evolution chains, type effectiveness, move pools, and geographic distribution
• Multi-source Hybrid Retrieval: Combines Milvus vector search, Neo4j graph retrieval, and BM25 keyword search
• Advanced RAG Strategies: Self-RAG/CRAG for reflection and retrieval correction, HyDE for hypothetical document embeddings, Query Decomposition for complex questions, Reranker for result re-ranking
• GraphRAG (LightRAG): Integrated HKU-DS LightRAG for enhanced graph construction and retrieval
• Rule-first Routing: Deterministic data (e.g., base stats) returned directly via rules without LLM calls

Multi-Agent Architecture#

Supervisor-Workers pattern implementation:

• Supervisor Agent: Intent recognition and task distribution
• RAG Worker: Document and knowledge base retrieval
• Web Worker: Tavily web search
• Graph Worker: Neo4j Cypher graph queries
• Stats Worker: Numerical statistics analysis
• MCP Worker: External tools and map services

Knowledge Processing#

• Domain Model: Fine-tuned Pokémon domain model based on Qwen2.5-14B-Instruct (INT4 quantization)
• Entity Recognition (NER): Triple guarantee with RoBERTa + TF-IDF + Rule matching
• Document Parsing: DeepDoc (from RAGflow) supporting complex layouts like PDF/Word
• Knowledge Base Workbench: Batch import, chunk management, retrieval testing

UI and Interaction#

• Rotom-Dex OS Theme: Authentic Pokémon game-style UI with dark mode support
• Voice Input: Integrated Alibaba DAMO Academy FunASR streaming speech recognition
• Map Visualization: MCP service for Pokémon geographic distribution display
• Graph Visualization: Neo4j relationship network visualization

Performance Optimization#

• Semantic Cache for reducing redundant computation
• Speculative RAG acceleration
• SQLite persistence for dialogue state and embedding cache

Deployment Guide#

Prerequisites#

• Docker & Docker Compose
• Node.js ≥ 18 (frontend local development)
• Python ≥ 3.11 (backend local development)

Quick Start#

# Clone repository
git clone https://github.com/skygazer42/pokemon-chat.git
cd pokemon-chat

# Configure environment
cp .env.example .env
# Edit .env and add LLM API Key (e.g., SILICONFLOW_API_KEY)

# Start core services
cd docker
docker compose up -d --build

# Start full infrastructure (with Neo4j/MySQL/Milvus)
docker compose --profile infra up -d --build

# Optional: Enable MCP service
docker compose --profile infra --profile mcp up -d --build

# Optional: Enable ASR service
docker compose --profile infra --profile asr up -d --build

Access Points#

• Web UI: http://localhost:3100/
• API Docs: http://localhost:5050/api/docs
• Neo4j Browser: http://localhost:7474/

Data Initialization#

When started with --profile infra, the neo4j-bootstrap container automatically imports:

• resources/data/kg_data/entities.json
• resources/data/kg_data/relations.json

Map data import:

docker compose exec api python scripts/import_pokemon_map.py

Configuration#

The project uses root .env file for unified configuration management, automatically loaded by Docker Compose.

Key Feature Toggles#

Model Configuration#

• LLM: llm_api_key, llm_api_base, llm_model_name, llm_temperature, llm_max_tokens
• Embedding: embedding_provider, embedding_api_key, embedding_model_name, embedding_dimension
• Reranker: reranker_model_name, reranker_top_k, reranker_threshold

Database Configuration#

• Neo4j: neo4j_uri (e.g., bolt://neo4j:7687), neo4j_username, neo4j_password
• Milvus: milvus_uri (e.g., http://milvus:19530)
• MySQL: mysql_host, mysql_port, mysql_user, mysql_password, mysql_database

Core Workflow#

1. Input Processing: Voice/Text → ASR (FunASR) → Text → NER entity extraction
2. Intent Routing: Supervisor Agent analyzes intent, decides direct rule-based response or worker dispatch
3. Retrieval Execution:
   • RAG Worker → Vector search + reranking
   • Graph Worker → Neo4j Cypher query
   • Web Worker → Tavily search
4. Generation & Synthesis: Worker results aggregation → LLM generates final response → Frontend rendering

Extension and Migration#

The architecture supports rapid migration to other vertical domains:

1. Replace knowledge source data (entities.json, relations.json)
2. Adjust Neo4j graph schema
3. Update domain-specific configurations and prompts
4. Optional: Fine-tune domain-specific model

Applicable scenarios: Financial Q&A, legal consulting, government services, and other domain applications requiring precise knowledge retrieval and relationship reasoning.