IHKIP — Intelligent Health Knowledge & Information Platform

Executive Overview

IHKIP is an ongoing research and engineering initiative exploring verifiable health knowledge retrieval. By organizing structured clinical guidance into high-dimensional vector embeddings, it enables rapid semantic search across medical documentation.

The Problem & Real-World Context

Healthcare professionals and organizations frequently struggle to quickly query dense clinical guidelines. Cloud-hosted LLMs introduce severe data privacy concerns, while keyword search fails on synonyms and medical phrasing.

System Design & Architecture

Structured ingestion pipeline chunking clinical guidance, indexing vector embeddings into PostgreSQL with pgvector, and orchestrating local quantized models (Qwen) with strict citation verification.

[Client Request / HTTPS] ────► [Reverse Proxy (Nginx/SSL)] ────► [Django Application Layer]
                                                                        │
        ┌───────────────────────────────────────────────────────────────┴─────────────────────────────┐
        ▼                                                                                             ▼
[Relational & Domain Engine]                                                              [Vector & RAG Pipeline]
  • PostgreSQL Storage                                                                      • Document Parsing & Chunking
  • Transaction Integrity & RBAC                                                           • pgvector Similarity Search (Top-K)
  • Automated Encrypted Backups                                                             • LLM Inference & Citation Engine
            

Technology Stack & Tooling

Implementation Details

Python/Django orchestration service, pgvector cosine distance queries, HNSW index optimization, and local inference server on Linux VPS.

Engineering Challenges & Key Solutions

Mitigated clinical hallucinations by requiring all generated insights to cite exact source chunks and rejecting unsupported claims.

Security, Access Control & Data Governance

Self-hosted on Linux VPS with zero external API calls to safeguard health knowledge sovereignty.

Outcomes & Verified Results

Sub-second retrieval across comprehensive guidelines with grounded citation accuracy. Ongoing active prototype.

Engineering Takeaways & Lessons Learned

Relational databases augmented with vector extensions provide better transactional consistency than detached vector stores.

Future Roadmap & Iterations

Multi-lingual clinical terminology indexing (Kinyarwanda & French) and automated guideline update diffs.