KNOWLEDGE ARCHITECTURE
Knowledge Base & Grounded RAG Indexing
How Crescelo reads your documentation, creates vector embeddings, and ensures that answers cite your exact material without inventing facts.
1. Supported File Types & Live Web Scraping
You can train your assistant on:
- • PDF & Word Documents: Price sheets, employee handbooks, service descriptions, and policy files.
- • CSV & Spreadsheets: Product catalogs, inventory rates, and structured FAQs.
- • Website URLs: Enter your domain URL. Crescelo automatically fetches and strips HTML boilerplate, indexing only high-value content.
2. Hybrid Dense Vector & BM25 Keyword Search
Crescelo utilizes local ONNX embeddings (Xenova/bge-small-en-v1.5) combined with BM25 lexical keyword ranking through Reciprocal Rank Fusion (RRF). If a visitor asks a colloquial question like “how much does it set me back?”, the vector engine recognizes that “set me back” corresponds semantically to your “Pricing” passage.
3. The Anti-Hallucination Guardrail
If a question cannot be verified against any passage in your knowledge base with sufficient similarity, Crescelo explicitly states that it does not have that information, rather than fabricating a confident but false answer.
4. Unanswered Query Diagnostics
Every time the assistant cannot find an answer, the question is logged in your CRM dashboard under Unanswered Queries. This allows your team to see exactly what prospective customers are asking for and upload a 1-sentence note to fill the gap permanently.