CiteGuard RAG - Sell Project
Listed November 15, 2026
About CiteGuard RAG
Production-ready, anti-hallucination RAG chatbot engine
Overview CiteGuard RAG is a complete, production-oriented Retrieval-Augmented Generation (RAG) system built specifically for strict policy, compliance, and academic/organizational handbooks—where standard LLM chatbots fail by hallucinating numbers, missing cross-references, or leaking sensitive inputs. I originally built and battle-tested this architecture on real university regulatory documents (multi-chapter handbooks with complex GPA thresholds, credit rules, and cross-referenced clauses). Because the target institution did not move forward with adoption, I have decoupled the domain data and am offering the clean codebase for sale to a builder, agency, or SaaS founder who wants a rock-solid RAG foundation without spending months engineering it from scratch. -- Key Technical Features (What Makes It Different from Basic RAG) Structure-Aware Ingestion & Quarantine: Parses PDFs (including OCR via Tesseract + pdfplumber) by Chapter/Article/Clause rather than naive fixed-size chunking. Includes an automated data contract & quarantine pipeline to isolate noisy OCR/table chunks so the LLM never cites corrupted figures, plus SHA-256 manifest change detection (1-command KB update). 4-Layer Hybrid Retrieval + 1-Hop GraphRAG: Combines Parent-Child MaxSim (baai/bge-m3 on FAISS) with BM25Okapi + Calibrated RRF, an interaction reranker (handles decimal formatting variations and numeric-unit pairs), and 1-hop cross-reference expansion for linked clauses. Neuro-Symbolic Rule Engine: Runs deterministic math/threshold checks in Python before passing context to the LLM—eliminating arithmetic and comparison errors. Strict Grounding & Safe Abstention: Forces structured JSON output with verified sourceid citations (document name, clause, page). If evidence is missing or a citation fails verification, the system automatically abstains and provides human escalation contacts. Defense-in-Depth Privacy & Ops: 4-layer PII filter blocking IDs, phone numbers, and emails before leaving the UI, plus untrusted-data prompt wrapping. Multi-query decomposition + conversational memory (last 3 turns). LRU + TTL query cache (never caches error responses). Queue service (100 concurrent / 500 queued limit) with InMemory or Redis worker support. Automatic LLM fallback (configured for Gemma 4 26B via OpenRouter with automatic failover to Nemotron 30B on rate limits/network errors). Tested & Verified: Includes an automated test suite (146/147 tests passing; 1 skipped when local Redis is off), plus preflight & operational rehearsal checks. -- Tech Stack Language / Backend: Python, FAISS (IndexFlatIP), rankbm25, baai/bge-m3 embeddings PDF / OCR: pdfplumber, PyPDF2, Tesseract OCR LLM Layer: OpenRouter API (Gemma 4 26B A4B MoE + Nemotron-3-Nano-30B fallback) Queue / Cache: In-Memory Queue / Redis Worker, LRU+TTL Cache Frontend: Streamlit UI (with Reasoning Trace & Citation Inspector) -- Traction & Status Status: Pre-revenue, MVP/Pilot-ready codebase (verified via automated tests and pilot rehearsals). Why I'm selling: Built as a passion project for my university; they haven't pursued deployment, and I'd rather hand this clean engineering work off to someone who can commercialize it as a B2B SaaS, agency boilerplate, or internal enterprise tool. -- Monetization Potential B2B Niche SaaS / White-Label: Package with a web widget/Zalo/Messenger integration for private academies, clinics, law firms, HR departments, or customer support teams needing strictly cited answers. Agency Boilerplate: Use it as a repeatable core engine to deliver custom enterprise RAG deployments at higher margins than no-code tools. -- What's Included in the Sale 1. Full clean Python source code (modular 4-layer architecture + Streamlit UI). 2. Automated PDF ingestion & 1-command KB update scripts (updateknowledgebase.py). 3. Full test suite (146+ tests) & preflight/rehearsal scripts. 4. Architecture diagrams, flowcharts, and setup documentation (with clean sample PDF data—no private institutional data included). 5. Post-sale support: Happy to provide a live walkthrough/handover call to help you run and customize the system.
Project details
Website: http://192.168.200.37:8501
For sale
Asking price: $250 USD
What's included
Source code
Reason for selling
Hiện tại tôi đang cần tiền và muốn bán dự án này.
Project metrics
Average Monthly Visitors: Pre-launch
Average Monthly Revenue: Undisclosed
Average Monthly Downloads: Pre-launch / New project