Core Concepts
Knowledge
Vector embeddings, multi-source ingestion, hybrid search, and semantic grounding.
Overview
The Knowledge Engine provides your assistant with factual, proprietary business context. It eliminates hallucinations by converting your documents, websites, and code repositories into searchable vector embeddings.
How Knowledge Works
When a visitor asks a question:
- Embedding Generation: The query is converted into a 1536-dimensional semantic vector.
- Hybrid Retrieval: Asiyst combines dense vector similarity (cosine distance) with keyword BM25 search across your project's vector store.
- Context Injection: The top 3–5 most relevant text chunks are formatted into the LLM system prompt along with verified source citations.
- Answer Synthesis: The model produces an accurate answer referencing only the supplied context.
User Query ──► Vectorization ──► Hybrid Search (Dense + BM25) ──► Top Chunks ──► LLM ContextSupported Source Types
- Website & Sitemaps: Crawl public URLs and XML sitemaps automatically.
- Uploaded Documents: PDF, Markdown (
.md,.mdx), Plain Text (.txt), and Word (.docx).
- GitHub Repositories: Read-only repository access for codebase, README, and documentation indexing.
- Manual Instructions: Structured Q&A pairs and business rules written directly in the dashboard.
Security & Redaction
During ingestion, the Asiyst parser automatically redacts:
- Secrets, private API tokens, and passwords matching standard regex patterns.
- PII (emails, phone numbers, credit card numbers) unless explicitly marked for retention.
Read more in Knowledge Base Overview.