12-stage RAG-based grounding pipeline + structured citation system. Every response is grounded in tenant content - not fine-tuned. Visitors see citation snippets per source. Confidence-scored. Audit-trail logged. Continuous eval feedback. Pairs with the architecture overview, answer-quality page and AI ethics statement.
Documents (PDF, DOCX, scraped HTML, FAQ pairs, manually-authored articles) chunked into 200-500 token segments with 50-token overlap. Each chunk gets metadata: ClientId (tenant partition), sourceUrl, chunkIndex, lastUpdated.
Technical: Chunking algorithm respects document structure (paragraph boundaries) when possible. Overlap prevents context-loss across chunk boundaries.
Our embedding model (1536 dimensions) converts each chunk into a semantic vector. Embeddings are stored in our vector store with the chunk's metadata.
Technical: Same embedding model used for queries - ensures cosine-similarity comparisons are semantically valid.
When a visitor submits a message, the bot embeds the query using the same embedding model. Query embedding is filtered against ClientId before retrieval - cross-tenant contamination is structurally impossible.
Technical: Repository pattern enforces ClientId filter at compile time via static analyzer rule (SLATECH001).
The vector store returns the top-K (default 10) chunks with highest cosine similarity to the query. Default ScoreThreshold = 0.5 filters out low-relevance chunks.
Technical: TopK clamped to [1, 20] per per-tenant configuration. Below-threshold queries route to a "no relevant content" fallback rather than hallucinating.
Retrieved chunks + system prompt + conversation history pass to the LLM. System prompt explicitly instructs the LLM: "Answer only from the provided context. If the answer is not in the context, say so."
Technical: Token budget enforced (default 4000 tokens of context); if budget exceeds, low-score chunks are dropped first.
A leading large language model (default) or a tenant-configured LLM generates the response, grounded in the retrieved context. Temperature default 0.3 for grounded customer-facing answers.
Technical: Per-tenant LLM provider abstraction allows swapping language-model providers without customer-side migration.
The response includes a structured citation list: { sourceUrl, snippet, score } per retrieved chunk. Snippet is the actual quoted text - not just the URL.
Technical: BuildSnippet helper in QueryRequest record extracts the relevant 200-character span from the chunk.
Server-Sent Events transport. First event is sources-early - emits citation metadata before LLM streaming starts. Widget renders "according to" hover-card while the answer is still streaming.
Technical: Cuts perceived latency by ~70% vs synchronous response. Enables AI scrapers to extract grounded quotes from the response.
Answers built on retrieved content are scored by a secondary call against three axes: factuality, hallucination and confidence. Scores surface in the admin Inbox.
Technical: Confidence score below 0.5 typically triggers a human-handoff fallback rather than a guessed answer.
When confidence is low OR query is identified as high-risk (clinical advice, legal position, regulatory question) - bot routes to a "human will follow up" pattern. The visitor receives an acknowledgement + a follow-up channel.
Technical: Per-vertical instructions tuned for each industry. Med is configured to refer diagnosis-adjacent questions to a clinician rather than answer them; Legal refers substantive legal questions to an attorney. A separate compliance check reads answers after they are sent and flags medical-advice and diagnosis claims in the admin Inbox, so the two work together: instruction first, detection behind it.
Every response logged with full context: input query, retrieved chunks with scores, system prompt, LLM model used, generated response, citation snippets, confidence scores.
Technical: Audit logs retained 13 months. Per-tenant audit log exportable on Enterprise tier.
Factuality and unsupported-statement flags are recorded per answer and can be tracked over time for a site. There is no nightly benchmark run: the sealed per-vertical test set described here never existed.
Technical: How answer quality is measured, and what a publishable comparison would need, is at /en/eval/.
Answers built on retrieved content are scored for factuality, and the results are visible to you for your own site. There is no published benchmark yet - see /en/eval/.