RAG Problem Tree Generator
AI-Powered Regional Policy & Causality Decomposition
- Role
- Creator & AI/Systems Experimenter
- Period
- 2024 - 2025
- Organisation
- Independent
Overview
An experimental AI system that parses regional planning documents and socio-economic datasets to generate verified causal problem trees (Masalah → Penyebab → Intervensi) using retrieval-augmented generation.
Context
Drafting policy problem trees typically requires weeks of manual document cross-referencing. This experiment tests whether local LLMs combined with vector embeddings can accurately construct valid causal graphs from unstructured government reports.
Problem
Traditional LLMs hallucinate non-existent regional regulations and invent unsubstantiated causes when asked to generate policy recommendations directly.
Approach
Architected a grounded RAG pipeline using ChromaDB vector stores and local Ollama embeddings to constrain the model strictly to verified regulatory texts and statistical datasets.
Flow
- Data
Regional Policy Data
Regional documents, RKPD reports, and statistical datasets
- Retrieval
Information Retrieval
Dense semantic search via ChromaDB vector embeddings
- RAG
AI Synthesis Pipeline
LangChain orchestration with domain-constrained prompt templates
- Problem Tree
Structured Output
Hierarchical graph: Masalah → Penyebab → Opsi Intervensi
Technical challenges
- Enforcing strict hierarchical schema output (valid JSON DAG) from language models without syntax breakage.
- Optimizing local embedding and inference speeds using Ollama on developer workstation environments.
Engineering decisions
- Implemented strict Pydantic schema validation wrappers to ensure generated problem trees conform to deterministic front-end graph nodes.
- Used hybrid chunking (document headers + semantic paragraphs) to preserve regional context during vector retrieval.
What shipped
- Grounded RAG architecture using ChromaDB and Ollama
- Deterministic schema parsing into interactive problem-tree graphs
- Hands-on technical exploration bridging AI capabilities with real-world planning needs
Stack
FastAPI · LangChain · ChromaDB · Ollama · Vector Embeddings · Document Processing