TikTok TechJam 2025 – GeoFlag

Feature → Laws Compliance Reasoning Pipeline

Problem

From Guesswork to Governance: Automating Geo-Regulation with LLM

Members

1. Daniel Kwan
2. Ivan Jerrick Koh
3. Shaun Tan Shu Ren
4. Gordon Hong Jia Jie

Overview

GeoRag is a compliance reasoning tool, accessed via Streamlit.
It takes product artifacts such as PRDs as input, analyzes them for potential compliance signals, retrieves relevant legal clauses from a Pinecone vector index, and runs a geo-specific compliance reasoner to determine whether implementation needs to vary across jurisdictions.

Installation

1. Clone repo

git clone https://github.com/danielkwan2004/TechJam2025
cd TechJam2025

2. Create a virtual environment

We recommend Python 3.13.

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

3. Configure environment variables and run

Copy .env.example → .env and set:

OPENAI_API_KEY=yourKeyHere
GOOGLE_API_KEY=yourKeyHere

Run:

cd langchain && docker compose up -d
streamlit run app.py

Tech Stack

Tools Used

• LLM Application Framework: LangChain and LangSmith
• Embedding Model: sentence-transformers/all-mpnet-base-v2
• UI Framework: Streamlit
• Databases: Pinecone

APIs Used

• OpenAI

Description

Pinecone Retriever

GeoRag uses Pinecone for vector search. Laws/clauses are pre-indexed into Pinecone with metadata:

{
  "type": "...",
  "article_number": "...",
  "recital_number": "...",
  "title": "..",
  "clauses": [
    {
      "clause_id": "...",
      "clause_text": "....",
      "signals": [...]
    }
  ]
}

• Pinecone is run locally via Docker.
• During retrieval:
  • Semantic search (HF embeddings like sentence-transformers/all-MiniLM-L6-v2)
  • Signal-filtered search (boosting candidates tagged with extracted signals)
  • Fusion (reciprocal rank + overlap bonus)
  • Top-K results returned to the reasoner

Signal Extractor

Abbreviation Handling

• Abbreviations are stored in a Supabase table:

abbreviations(term, explanation)

• Includes those provided in Data Set → Terminology Table.
• At runtime, abbreviation_helper.py queries Supabase for all known terms.
• If a term appears in PRD text, it is included in the LLM context.

Example:

from abbreviation_helper import retrieve_abbreviations

retrieve_abbreviations("Upload limits apply for IMT accounts.")
# → [{'term': 'IMT', 'explanation': 'In-Market Testing'}, ...]

Add new abbreviations via Streamlit frontend:

from abbreviation_helper import add_abbreviation

add_abbreviation("BB", "Behavioral Benchmarking")

Law Cards

GeoRag relies on a structured JSON knowledge base called law_cards.json.
Each card represents a regulation, act, or statute, defining keywords, triggers, and signals that map product features to legal obligations.

Example Structure:

{
  "version": "1.1",
  "laws": [
    {
      "id": "US_ALL_2258A",
      "name": "U.S.Code 2258A - NCMEC reporting requirements of providers",
      "region": "US_all",
      "category": [...],
      "triggers": {
        "keywords": [...],
        "signals": [...],
        "must_all": [...],
        "must_any": [
          "ncmec_reference",
          "cyphertipline_reference",
          "include_visual_depictions",
          "include_complete_communication"
        ],
        "negations": [...]
      },
      "reason_templates": [...],
      "priority": "high",
      "last_updated": "2025-08-26"
    }
  ]
}

Pipeline

1. Signal Extraction (LLM)

• PRD text analyzed for keyword matches
• Corresponding signals suggested
• must_all / must_any logic ensures valid triggers
• negations reduce false positives
• LLM corroborates signals (with abbreviations substituted)
• Outputs signals with fields:

{
  "law": "...",
  "signal": "...",
  "reason": "...",
  "confidence": "..."
}

2. Retrieval (Pinecone)

• Extracted signals query Pinecone index of law clauses
• Clauses tagged with signals are boosted in ranking

3. Reasoning (LLM)

• Feature + retrieved clauses fed into reasoner
• Outputs final decision:

{
  "geo_specific_required": true,
  "reasoning": "...",
  "related_regulations": [
    {"id": "13-63-103(4)", "title": null, "snippet": ""}
  ]
}