multimodal-evidence-sdk 0.1.0

A Python SDK for verifying, ranking, and reviewing multi-modal damage claims using Gemini vision models.

Multimodal Evidence SDK

multimodal-evidence-sdk is a Python SDK for verifying, ranking, and reviewing multi-modal damage claims using Gemini vision models. It provides structural validation and decision-making logic originally developed for the HackerRank Orchestrate Evidence Review platform.

---

Installation

Install the package locally in editable mode:

pip install -e .

---

SDK Usage

from multimodal_evidence import (
    retrieve_evidence,
    rank_evidence,
    verify_claim
)

# 1. Verify a claim statement factually (Text-only)
result = verify_claim(
    claim_text="India launched Chandrayaan-3 in 2023"
)
print(result["claim_status"])  # -> "supported"

# 2. Verify a damage claim with image evidence (Multimodal)
claim_result = verify_claim(
    claim_text="The car rear bumper has a major dent",
    images=["path/to/img_1.jpg"],
    claim_object="car",
    history={
        "user_id": "usr_123",
        "past_claim_count": 0,
        "accept_claim": 0,
        "manual_review_claim": 0,
        "rejected_claim": 0,
        "last_90_days_claim_count": 0,
        "history_flags": "none",
        "history_summary": ""
    }
)
print(claim_result["claim_status"])

---

Command Line Interface (CLI)

The SDK registers the evidence executable.

1. Verify a Statement

evidence verify "The Earth has two moons"

For damage claims with images:

evidence verify "Rear bumper dent" --images "dataset/images/sample/case_001/img_1.jpg" --object car

2. Search Factual Details or Requirements

evidence search "Chandrayaan-3 launch"

evidence search car

3. Rank Evidence from JSON Payload

evidence rank evidence.json

---

Development & Evaluation

If you are participating in the HackerRank Orchestrate hackathon, you can run the evaluation metrics and pipeline using the scripts in code/:

1. Install Dependencies

cd code/
pip install -r requirements.txt

2. Set API Key

Copy the template .env.example to .env and insert your API key:

# code/.env
GOOGLE_API_KEY=your_gemini_api_key_here

3. Run Pipeline on Test Data

cd code/
python main.py

4. Run Strategy Evaluation

cd code/
python evaluation/main.py --compare

---

Project Structure

• multimodal_evidence/: The SDK package directory.
  • models/: Pydantic schemas and enums.
  • retrieval/: Search and requirements indexing.
  • ranking/: Guardrails and risk flag merging.
  • verification/: Claims and factual verifications.
  • multimodal/: Gemini client wrapper and batch pipeline runner.
• code/: The hackathon pipeline interface (fully refactored to consume the SDK).
• dataset/: Claims databases and visual evidence.