Under the hood
Most tools won’t tell you what’s running behind the answers. Here is the complete picture: the two models involved, why they were chosen, and the one case where SiteMind answers your visitor without calling a model at all.
The models
One turns your content and your visitors’ questions into something searchable. The other reads what was found and writes the reply.
Writing the answer
Gemini Flash-Lite
Retrieval has already found the right passages, so the model’s job is to read and summarise them — not to recall facts. That job does not need a frontier model, and a faster one means a visibly quicker answer.
Understanding your content
gemini-embedding-001, 768 dimensions
Turns every chunk of your content, and every visitor question, into a vector so the right passage can be found by meaning rather than keyword match.
How it works
Five stages, in order. Note how far down the list the language model appears — and that it can be skipped entirely.
The question is converted into a vector using the embedding model, as a query rather than a document — the two are embedded differently, and mixing them degrades search quality.
That vector is compared against every chunk of your indexed content to find the passages that genuinely match the meaning of the question.
The assistant returns a fixed "I couldn’t find that information on the website" and stops. No language model is invoked, so there is nothing available to invent an answer with.
The retrieved passages are assembled into a prompt with instructions to answer only from that material, along with the last few turns of the conversation for context.
Only now does the chat model run. It reads the supplied passages and writes an answer from them, which is streamed back with citations linking to the pages the content came from.
How we think about it
“Which LLM does it use?” is the question everyone asks. It’s rarely the one that decides whether the answers are any good.
If the right passage reaches the model, a small model answers correctly. If it doesn’t, the largest model available will still get it wrong — it just sounds more convincing while doing so. That is why we invest in search quality rather than in a bigger model, and why "which LLM does it use" is a less useful question than it appears.
Chat and embeddings both sit behind provider interfaces, so the underlying model can be changed without touching the knowledge pipeline. Each stored vector records the provider, model and dimensions that produced it, so a migration stays traceable rather than silently mixing incompatible embeddings.
Most products send every question to the model no matter what. SiteMind treats the model call as the last step and skips it entirely when retrieval finds nothing — which is what makes an honest refusal structural rather than a matter of prompt wording.
Your content is used to answer your own visitors and nothing else. It is not used to train foundation models and it is not sold. See the security page for the full data-handling detail.
Limitations
The honest constraints, so you can rule us out quickly if one of these is a dealbreaker.
SiteMind runs one configured chat model for everyone rather than offering a model selector. The architecture supports swapping it, but choosing per workspace is not something you can do from the dashboard. If you have a specific requirement here, talk to us.
The assistant cannot answer from the model’s general world knowledge, by design. Anything not in your crawled pages, uploaded documents or connected sources will be declined rather than guessed at.
SiteMind is a hosted product and calls the model provider with our own credentials. There is currently no way to supply your own API key or run the models in your own environment.
FAQ
SiteMind generates answers with Google’s Gemini Flash-Lite model and builds its search index with Gemini’s embedding model (gemini-embedding-001) at 768 dimensions. Both sit behind provider abstractions, so the underlying model can be swapped without changing how the product works.
Because the model is not doing the hard part. In a retrieval-based system, search finds the relevant passages and the model only has to read and summarise them. A larger model would make that step slower and more expensive without making it more accurate — the accuracy comes from retrieval. Where a bigger model genuinely helps is open-ended reasoning, which is deliberately not what this product does.
No. SiteMind currently runs on Google’s Gemini models for both chat and embeddings. The provider layer is built to be swappable, so this could change, but any change would be reflected on this page.
No. Your content is indexed so it can be retrieved and used to answer your own visitors’ questions. It is not used to train foundation models and it is not sold or shared. Deleting a workspace removes your data within 30 days.
Not from the dashboard today. Every workspace runs the same configured chat model. The system is architected so the model can be swapped centrally, and a per-workspace choice is possible in principle, but it is not a feature you can use right now. Enterprise customers with a specific model or region requirement should get in touch.
Answer generation is the only stage that depends on the provider, so an outage affects new answers rather than your indexed knowledge, which lives in our own database. Requests are spread across a pool of API keys with automatic failover between them, which covers per-key quota exhaustion — though it cannot help with a full provider-side outage.
Point SiteMind at your website, then ask it something your content doesn’t cover. A trustworthy assistant says so. Free for 3 days, no card required.
Related: Features · What is RAG? · Why chatbots hallucinate