A generic AI chatbot guesses. A grounded one retrieves
Any LLM can answer a question about your product with confidence — the problem is when it's wrong. JestBot retrieves from your own docs first, and only answers from what it actually found.
No card required · Free plan included
Cosine similarity
Retrieval method
6
Model providers
< 2 min
Setup time
No card needed
Free plan
Why 'just add ChatGPT' doesn't work
What goes wrong with an ungrounded AI chatbot
A chatbot that answers from general training data instead of your actual product will eventually say something confidently wrong — and a wrong answer about your product is worse than no answer.
It doesn't know what you actually shipped
General training data has no idea about your specific plans, pricing, or last week's feature release — so it fills the gap with a plausible-sounding guess.
You can't see why it said what it said
Without visible retrieval, a wrong answer is something you distrust, not something you can debug and fix.
It has nowhere to go when it's stuck
A chatbot with no handover path either keeps guessing or leaves the customer stranded — neither builds trust.
How a grounded AI chatbot actually works
01
Retrieve
The question is matched against your own documents using cosine-similarity embedding search — not the model's memory.
02
Generate
Only the most relevant retrieved chunks are handed to the model as context for its answer, visible with a similarity score in Test Chat.
03
Escalate
If a conversation needs judgment or empathy, one click hands it to a real agent with the full context intact.
What makes it grounded
Retrieval-augmented generation, without the infrastructure
RAG usually means running your own vector database and prompt pipeline. JestBot handles ingestion, retrieval, and generation end to end.
Ingest your product's actual knowledge
Paste raw text, point it at a URL, or upload a file — PDF, DOCX, TXT, CSV, or Markdown. Every source is chunked into overlapping ~1500-character windows so answers stay grounded in context.
- Text, URL, or file ingestion
- Automatic chunking with overlap
- No separate vector DB to run
See exactly what it retrieved
Every reply is built from the chunks the bot actually retrieved for that question — visible alongside the response in Test Chat with a similarity score — instead of the model improvising from training data.
- Cosine-similarity retrieval
- Retrieved chunks visible for debugging
- Streaming responses over SSE
Bring your own model
Start on the free default and switch any bot to your own OpenAI, Anthropic, Gemini, Groq, or Mistral key the moment you need a stronger model — encrypted at rest, never logged in plaintext.
- Free tier included by default
- Per-bot provider & model selection
- AES-256-GCM encrypted keys
Human handover when retrieval isn't enough
Some questions need judgment, not just information. Set a message threshold and a real agent can take over the live conversation instantly, with zero context lost.
- Configurable handover threshold
- Assign to agents or teams
- Real-time handoff, no message loss
Where this matters most
Grounding matters wherever a wrong answer costs you
The gap between a generic chatbot and a grounded one shows up fastest in these spots.
Plan & pricing accuracy
Get plan details right every time — a wrong pricing answer is a support ticket and a trust problem.
Developer & API questions
Ground answers in your actual API reference instead of a plausible-sounding but outdated guess.
Onboarding accuracy
Walk new users through setup steps that match your current product, not a generic description of 'how SaaS onboarding usually works.'
Answers that track your changelog
Re-ingest your docs after a release so answers reflect what shipped, not what used to be true.
Consistent answers across languages
The same grounded knowledge base replies correctly whether a visitor writes in English, Hindi, or Spanish.
Answers you can actually audit
Retrieved chunks and similarity scores mean a wrong answer is debuggable, not just a black box you distrust.
“Being able to see the retrieved chunks in Test Chat is what sold me — I could actually debug a bad answer instead of guessing whether the model was making things up.”
Marcus T.
Developer relations
Questions people actually ask
JestBot — AI Chatbot for SaaS