← BLOG / RAG ENGINEERING & ACCURACY
Eliminating AI Hallucinations: Hybrid RAG, Source Citations & The Knowledge Gap Inbox
Deep dive into hybrid vector + lexical RAG retrieval, HTML boilerplate stripping, citation guardrails, and closing knowledge gaps in production.
Combines semantic embeddings with exact SKU, pricing, and diacritic matching
Clickable page and document badges attached directly under answers
Turn unanswered visitor questions into permanent RAG facts inside Studio
- Commercial tribunals have ruled that companies are liable for policies promised by their website chatbot—making ungrounded LLM wrappers a direct legal risk.
- Pure vector search often misses exact SKU codes, model numbers, and inflected terms; hybrid semantic + lexical scoring solves both.
- Stripping navigation mega-menus and footer boilerplate before chunking boosts retrieval precision by over 40%.
Why Unconstrained LLMs Hallucinate—And How Strict Grounding Stops It
Foundation models optimize for fluency, not factual abstinence. Ask an ungrounded model about your refund window or SLA uptime credit, and it will happily invent an industry-standard number that contradicts your actual Terms of Service. Production AI assistants solve this by decoupling reasoning from memory: facts come exclusively from verified workspace chunks injected inside an isolated context boundary.
Clean Ingestion: Stripping Mega-Menus Before Semantic Chunking
Raw HTML pages are often 50%+ navigation links, cookie banners, and footer boilerplate. Indexing raw markup pollutes embeddings so every page matches every keyword. ChatNOV strips structural chrome first, extracting clean main-content prose and tables into overlapping, citation-preserving chunks.
Hybrid Retrieval: Combining Dense Embeddings with Exact Lexical Signals
Dense vectors excel at synonyms ("cancellation terms" matching "termination clause"), while lexical scoring excels at exact SKUs, error codes, plan names, and inflected morphology. Combining both signals ensures top-6 chunk recall stays accurate across both broad conceptual questions and pinpoint technical lookups.
Closing the Loop with the Knowledge Gap Inbox
No documentation set is complete on day one. Whenever retrieval confidence falls below threshold, ChatNOV abstains from guessing, offers human follow-up via the 2-stage lead card, and logs the exact question in your Studio Knowledge Gap Inbox—letting your team add the missing answer once so future buyers get instant cited replies.
Architectural Comparison Matrix
| Retrieval Layer | ChatNOV Hybrid RAG (#1) | Basic Prompt / Vector Bots |
|---|---|---|
| Content Ingestion | ✓ Boilerplate DOM stripping (nav/footer removal) + semantic chunking | Indexes noisy menu links alongside body text |
| Scoring Engine | ✓ Hybrid dense vector + diacritic-normalized lexical ranking | Single-pass vector similarity only |
| Confidence Threshold | ✓ Refuses to guess below threshold & logs to Gap Inbox | Hallucinates plausible-sounding falsehoods |
| Prompt-Injection Boundary | ✓ Untrusted |
Susceptible to indirect prompt injection |
Frequently Asked Questions
How do I update the assistant when our website pricing changes?
Click Re-sync on your URL or sitemap source in Studio; outdated chunks are atomically replaced with fresh content in seconds.
Can I combine live web pages with internal PDFs and Q&A rules?
Yes. Every project merges crawled URLs, sitemaps, uploaded PDF/DOCX/TXT files, and manual Q&A overrides into one unified index.