GroundTruth

Overview

GroundTruth is a zero-hallucination enterprise knowledge agent built for the Gemini Live Agent Challenge 2026. It delivers accurate, citation-grounded responses through real-time multimodal interaction using voice, camera vision, and text input.

The Problem

Enterprise AI agents powered by LLMs frequently hallucinate, producing confident but factually incorrect answers. In regulated industries like finance, healthcare, and legal, these hallucinations can have serious consequences. Existing AI chatbots lack robust grounding and citation mechanisms.

The Solution

GroundTruth implements a 5-layer grounding architecture that ensures every response is:

• Cited: Every claim references a specific source
• Validated: Post-response checker verifies all Source N references exist
• Transparent: Full audit trail of every query, citation, and validation
• Secure: Prompt injection defense with content treated as data, not instructions

Key Features

Multimodal Interaction

• Voice: Real-time streaming audio via Gemini 2.5 Flash Native Audio Live API
• Vision: Live camera feed analysis - point at any document for instant answers
• Text: Traditional text-based query interface

Zero-Hallucination Architecture

• BM25 retrieval for accurate document chunk matching
• 5-layer grounding pipeline with citation enforcement
• Post-response validation that confirms every citation is real

Enterprise-Ready

• Deployed on Google Cloud Run with Docker
• Infrastructure-as-Code via Terraform
• Full audit logging and transparency
• Professional Kore voice persona

How It Works

1. Document Ingestion: Upload PDF, DOCX, or TXT files. Documents are chunked and indexed with BM25 retrieval.
2. Query: Ask questions via voice, camera (show a document), or text.
3. Grounded Retrieval: The grounding engine retrieves relevant chunks from uploaded documents.
4. Response Generation: Gemini 2.5 Flash Native Audio generates responses with inline Source N citations.
5. Citation Validation: Post-response checker verifies every citation reference exists in the source material.
6. Audit Trail: Every interaction is logged for full transparency.

Tech Stack

Demo

See the full demo video showing voice interaction, camera-based document analysis, citation display, and hallucination resistance testing.

What I Learned

• Gemini Live API real-time audio streaming and session management
• Designing hallucination-resistant AI agent architectures
• Building multimodal frontends with Web Audio and MediaDevices APIs
• Deploying production-ready AI services on Google Cloud Run
• Implementing Infrastructure-as-Code with Terraform for reproducible deployments

Challenges

Gemini Python FastAPI Google Cloud Docker Terraform WebSockets HTML5

• Managing real-time WebSocket audio streaming with low latency
• Ensuring citation accuracy across different document formats
• Balancing response quality with strict grounding constraints
• Building a responsive UI that handles voice, vision, and text simultaneously

Built With

• cloud
• fastapi
• gemini-2.5-flash
• google
• python