Medical_Callbot_GraphRAG

Medical Callbot Assist (GraphRAG, VI-enabled)

Version 1: Medical Chatbot (GraphRAG, VI-enabled)

Overview

• A medical assistant chatbot that answers in Vietnamese with full diacritics and helps route users to the right department and schedule an appointment.
• Uses a hybrid GraphRAG pipeline: vector retrieval for unstructured knowledge + lightweight graph/CSV knowledge (departments, doctors, schedules, curated diseases) to ground responses and booking.
• Designed to generalize across common triage scenarios; avoids brittle, hard-coded case logic by combining heuristics, embeddings, and structured lookups.

Key Capabilities

• Vietnamese-first answering: forces clear, concise answers in Vietnamese; auto-rewrites non-VI generations.
• Graph + Vector Retrieval: merges graph facts (curated from medical_kg.csv and infor.md) with top‑K text chunks from embeddings.
• Symptom → Department routing: infers suspects (e.g., viêm ruột thừa, tiêu chảy cấp) and maps them to the most suitable department (Ngoại, Tiêu Hóa, Nhi, Sản...).
• Profile-aware booking: extracts basic patient profile (tuổi, giới tính, mang thai) from conversation and uses it to route (e.g., Nhi only when ≤12 tuổi) and to pick suitable doctors/slots.
• Structured booking: reads doctors, rooms, schedules, and channels from CSV to suggest options and can hold a provisional slot.
• Backend-flexible LLM: works with Ollama or Hugging Face pipelines; configurable model/temperature/top‑K.

How It Works (High Level)

1. Ingestion

   • Build sentence embeddings for documents into FAISS under embeddings/.
   • Maintain mapping of rows → ids and metadata (texts.jsonl, meta.jsonl, model.txt).

2. Retrieval

   • Vector search top‑K relevant chunks for the user query.
   • Graph facts inference via rule-based symptom matching (accent-insensitive) to produce concise facts and suspects.
   • Department routing with three layers: • Knowledge mapping from data/infor.md (curated diseases → department) • Heuristic keyword rules (e.g., đau quanh rốn → hố chậu phải → nghi viêm ruột thừa → Ngoại) • Embedding-based router tie-break (DepartmentRouter)

3. Conversation + Reasoning

   • Build a compact prompt with: system policy (Vietnamese only), Graph facts, retrieved context, recent hypothesis carry‑over, and patient profile summary.
   • Call LLM (Ollama preferred, HF fallback) and enforce Vietnamese output.

4. Booking Flow

   • Detect booking intent ("đặt lịch", "hẹn khám", ...).
   • Parse optional doctor name/weekday and select matching schedule; otherwise auto-pick department → doctors → earliest feasible slot.
   • Patient profile influences routing: e.g., ≤12 tuổi → Nhi; pregnant → Sản/Phụ sản.
   • Emit a clear Vietnamese confirmation (tạm giữ chỗ) with thời gian, phòng, bác sĩ, hotline; append profile summary.

Data & Knowledge Sources

Structured Data (CSV)

• data/departments.csv — departments (id, name, aliases, floors, notes)
• data/doctors.csv — doctors (specialization, titles, phones, department)
• data/schedules.csv — day-of-week/time slots, room ids
• data/rooms.csv — room labels per department
• data/booking_channels.csv — hotline, web, counter, app

Knowledge Graph & Curated Mappings

• data/medical_kg.csv — head/rel/tail triples used to enrich graph facts
• data/infor.md — curated disease → department lists (parsed to a mapping)

Medical Q&A Dataset

• ViMedAQA (Hugging Face)
  • A comprehensive Vietnamese medical question-answering dataset
  • Provides domain-specific medical knowledge in Vietnamese language
  • Used to enhance the vector retrieval component of GraphRAG
  • Helps ground responses with authentic Vietnamese medical terminology and common patient questions
  • Ingested into FAISS embeddings for semantic search during retrieval phase
  • Improves accuracy for symptom description understanding and medical advice generation

Integration Approach:

• ViMedAQA question-answer pairs are preprocessed and embedded alongside other medical documents
• During retrieval, relevant Q&A pairs are surfaced to provide contextually appropriate medical guidance
• Answers are adapted to fit the chatbot's structured response format (suspected diagnosis, self-care, red flags, booking)
• Helps bridge the gap between formal medical literature and conversational patient language

Generalization Strategy

• Accent-agnostic normalization for matching Vietnamese inputs reliably.
• Compose routing decisions from: (facts + suspects) ∪ (curated map) ∪ (router embeddings) to avoid single-point brittleness.
• Patient-profile extraction embedded into prompt and booking to reduce misrouting (e.g., pediatric vs adult appendicitis).

What You Can Expect

• Short, structured medical guidance (suspected diagnosis, safe self-care, red flags, booking suggestions) in Vietnamese.
• Evidence-citing prompt context (graph facts + retrieved snippets) to keep answers grounded.
• Deterministic, CSV-driven booking proposals that you can later wire to a real HIS system.

Demo Image:

Safety & Limitations

• Educational/triage assistant; not a substitute for a licensed clinician.
• Always advises escalation for red-flag symptoms (đau tăng, sốt cao, nôn liên tục, máu trong phân, mất nước nặng...).
• Coverage reflects the curated infor.md, ViMedAQA dataset, and available texts; expand these sources to broaden scope.

Roadmap

• Expand curated disease → department map, incl. pediatric vs adult nuances.
• Add lightweight forms to explicitly capture tuổi/giới tính/mang thai/thời điểm khởi phát.
• Optional appointment persistence to a DB or HIS API.
• Add evaluation harness of conversation transcripts for regression checks.
• Continuous expansion of ViMedAQA coverage and medical knowledge base.

Version 2: Callbot (...update soon...)

Contact (full code access)

• Le Doan Tho
• GitHub: https://github.com/tleeds1
• Email: thodoanle26@gmail.com