How ldlinglingling/rag-of-law works
Project Type and Purpose
• This is an ML-centric “full‐stack” service for legal question answering built on a Retrieval‐Augmented Fine-Tuning (RAFT) framework.
• It comprises offline training pipelines, an inference pipeline, a lightweight web service, and backing data stores (legal clause library, embeddings, BM25 index).
File Structure
Top‐level files:
• README.md → project overview, architecture summary
• fatiao.json → raw legal clause library
• training_data.json → Q&A training set
• training_data_sample.tsv.gz → sample training data
Training/inference scripts:
• data_proc.py → preprocessing pipeline (tokenization, indexing)
• bert_cls.py → BERT‐based two‐level classification model (12 categories + 115 subcategories)
• train_data_extractor.py → supervised fine-tuning script for answer extraction
• train_bi_encoder.py → contrastive training for semantic retrieval encoder
• extract_law_embs.py → generate & store law‐clause embeddings
• sample_run.py → simple end-to-end demo
• infer.py → core inference pipeline (classification → retrieval → generation)
• web_main.py → web server entrypoint (API endpoints)
README Insights
• The logical flow is:
User question
↓ Classification model (bert_cls.py)
↓ Law category → Semantic retrieval (BM25 + bi-encoder embeddings)
↓ LLM generation → Final answer
• Core components: classification, retrieval, generation.
• Data: legal clause library and Q&A dataset.
Main Components & Their Relationships
A. Frontend/API Layer
• web_main.py exposes REST or HTTP endpoints.
• Receives user questions; returns professional answers.
B. Classification Service
• bert_cls.py loaded by infer.py
• Determines broad/sub‐category
C. Retrieval Service
• BM25 index (from data_proc.py + inverted index)
• Vector index (embeddings from extract_law_embs.py)
• Bi‐encoder model (train_bi_encoder.py) for semantic scoring
D. Generation Service
• LLM (external API or local model) called by infer.py
• Uses retrieved clauses + question as prompt
E. Data Stores
• Raw clauses (fatiao.json)
• Precomputed embeddings (file‐based or vector DB)
• BM25 inverted index
• Training datasets
F. Offline Training Pipelines
• Preprocessing (data_proc.py) → cleaned JSON/TSV
• Classification training (bert_cls.py)
• Retrieval encoder training (train_bi_encoder.py)
• Fine-tuning generator (train_data_extractor.py)
Architectural Patterns & Design Principles
• Pipeline pattern: sequential stages (classify → retrieve → generate)
• Separation of concerns: offline training modules vs. online inference modules
• Modular ML microservices: each core function can be containerized
• Data‐driven: indexes and embeddings persisted externally for fast lookup
Essential Elements
• Clear component names