Half the calls ended in silence.

The agent said its line - "one moment, let me check the calendar" — and then did nothing. No error, no exception. The transcript looked perfect, which is exactly why it survived for weeks.

I found it by ignoring what the agent said and counting what it actually did: tool invocations in the call artifacts. On roughly half of all calls: zero.

The model was narrating the action instead of performing it. Two changes fixed it - a larger model, and moving the filler line to the platform's request-start hook, so speech and execution stopped sharing a path.

The lesson I keep re-learning with voice agents: transcripts lie. They show you what was said, not what happened. If you are debugging a live agent, count the tool calls.

Yes yes. You are on point

AI Assistant Using Your Business Knowledge Base — RAG on Your Documents

THE PROBLEM Q&A bots break down when knowledge lives in documents: a 100-page manual has no "questions" to match, it can't fit into a prompt, and generic chatbots hallucinate instead of admitting what they don't know.

THE SOLUTION A RAG (Retrieval-Augmented Generation) knowledge base: documents are split into meaningful chunks, embedded into a vector index, and the assistant answers from the right sections — by meaning, not keywords.

Any format as-is: PDF, DOCX, TXT, Markdown — 100+ pages is fine

Answers grounded in YOUR documents — it says "I don't have that information" rather than inventing

Source references — every answer shows which document and section it came from

Runs on your infrastructure — documents never leave your control

One command to re-index after updating documents — documented, no programmer needed

LIVE DEMO Ask it about my products, pricing, deployment and support — it answers from 6 real documents with sources: https://kb-assistant-demo.dvc-ai.workers.dev

The 'I don't have that information' line is the part most RAG builds skip, and it's the one that matters. How do you set the cutoff? On mine, a fixed similarity threshold broke once I filtered results by user role. Scores shifted and it refused questions it could answer.