API v2.0.0 · Operational

AI Agents That Know What They Don't Know

Build domain-specific AI agents from your documents. Every answer is grounded in your data, cited with inline references, and scored for hallucination risk using DeBERTa-v3 NLI.

📖 Explore API Docs 🚀 Launch App ⭐ GitHub

~1.2s Faithfulness Scoring

0.907 Guardrail F1 Score

781 Indexed Vector Chunks

5+ Groq Model Fallbacks

How It Works

The MEXAR RAG Pipeline

Every query goes through a 9-stage intelligent pipeline — from multimodal input all the way to a cited, faithfulness-verified response.

🎤 Multimodal Input Processing optional

Audio → Groq Whisper v3 STT · Images → Groq Vision · Video → OpenCV frame extraction → Vision

🔍 Intent & Prompt Analysis LLM

Parse query intent (factual / analytical / comparative), detect domain topic, optionally rewrite query for retrieval clarity.

🛡️ Domain Guardrail Check F1 = 0.9072

TF-IDF cosine similarity vs agent signature + spaCy NER Jaccard overlap. Threshold = 0.25. Out-of-domain queries rejected with explanation — no hallucination.

⚡ Hybrid Vector + Keyword Retrieval pgvector + BM25

Dense: FastEmbed bge-small-en-v1.5 (384-dim) cosine search via pgvector. Sparse: PostgreSQL tsvector BM25 full-text search. Fused via Reciprocal Rank Fusion (RRF): score = Σ 1/(rank + 60).

🎯 Cross-Encoder Reranking sentence-transformers

Re-scores top-20 retrieved candidates using a cross-encoder for precision. Selects final top-5 context chunks for answer generation.

🧠 LLM Answer Generation Groq

System prompt with retrieved context. Multi-model inference: llama-3.3-70b → llama-3.1-8b → mixtral-8x7b → gemma2-9b. Automatic quota fallback.

📎 Source Attribution citations

Match answer sentences to source chunks. Assign [1], [2], [3] inline reference markers. Track provenance per claim.

✅ DeBERTa-v3 Faithfulness Scoring NLI

Extract claims from answer. For each claim-chunk pair: NLI inference (entailment → faithful, contradiction → hallucinated). Batched via torch.inference_mode() — ~1.2s/query (50× speedup vs baseline).

🔬 Explainability Packaging transparent

Reasoning trace · Confidence breakdown · Sources cited · Guardrail decision log — all surfaced to the frontend UI panel.

Capabilities

Everything You Need

A complete RAG platform — from document ingestion to explainable, grounded answers.

🔍

Hybrid Search + RRF Fusion

Vector cosine (pgvector) and BM25 keyword search fused via Reciprocal Rank Fusion for optimal retrieval across all document types.

✅

Faithfulness Verification

DeBERTa-v3-base NLI model scores every answer claim against retrieved context. Quantified hallucination risk, not just vibes.

🛡️

Domain Guardrails

TF-IDF + spaCy NER Jaccard similarity prevents answering out-of-domain questions. Tuned to F1 = 0.9072 at threshold 0.25.

📎

Inline Source Citations

Every sentence references its source chunk with [1], [2] markers. Click any citation to see the exact source text and file name.

🗣️

Multimodal Input

Ask questions via audio (Groq Whisper), upload images for visual Q&A (Groq Vision), or extract info from video frames (OpenCV).

🧠

Explainability Panel

Full reasoning trace exposed in the UI: retrieval scores, reranker scores, confidence breakdown, guardrail decision, and sources cited.

🔊

Text-to-Speech

ElevenLabs API integration with per-sentence TTS playback. Falls back to Web Speech API automatically.

⚡

Real-time WebSocket Chat

Streaming responses via WebSocket with compilation progress tracking. No polling required.

📁

5 Document Formats

Upload PDF, DOCX, CSV, JSON, or TXT files. Semantic chunking preserves context boundaries for better retrieval quality.

REST API

API Endpoints

Full interactive documentation available at /docs

Method	Endpoint	Description
POST	/api/auth/register	Create a new user account
POST	/api/auth/login	Login and receive JWT bearer token
GET	/api/agents/	List all compiled agents for current user
POST	/api/agents/	Create a new agent
POST	/api/compile/	Start knowledge compilation from uploaded files
GET	/api/compile/{job_id}	Poll compilation job status + progress
POST	/api/chat/	Send a query to an agent (REST)
WS	/ws/chat/{agent_id}	Real-time streaming chat via WebSocket
GET	/api/health	Health check — returns API + Groq status

Reliability

Multi-Model Fallback Chain

Automatic failover across Groq models when rate limits are hit — zero downtime.

openai/gpt-oss-120b → quota →

llama-3.3-70b-versatile → quota →

llama-3.1-8b-instant → quota →

mixtral-8x7b-32768 → quota →

gemma2-9b-it

Getting Started

Quick Integration

Start querying your agent in three steps.

# 1. Register and login
POST /api/auth/register  { "email": "you@example.com", "password": "..." }
POST /api/auth/login     → { "access_token": "eyJ..." }

# 2. Compile an agent from your documents
POST /api/compile/  Authorization: Bearer {token}
     Files: report.pdf, research.docx
     → { "job_id": 42, "status": "compiling" }

# 3. Chat with your agent
POST /api/chat/  Authorization: Bearer {token}
     { "agent_id": 36, "message": "What are the key findings?" }
     → { "answer": "...[1][2]", "faithfulness": 0.87, "sources": [...] }
        