Keenious uses generative AI in two places: Chat, where you can ask communicate with the AI, and the short names shown for research areas on your results. Search matching itself uses embeddings (matching by meaning) and keyword ranking; search results are retrieved from the index, not generated by a language model.
The model behind Chat
Chat runs on Google's Gemini 3.5 Flash, selected for its document understanding and low latency. The short names shown for research areas are generated by a smaller, faster Gemini model.
Keenious does not offer a choice of model. In testing, allowing users to switch models caused confusion and, in some cases, more frequent hallucination, so a single tested model is used instead.
Hallucinations
Chat displays the reminder "AI can make mistakes. Please verify information." Each release is checked against an internal set of test scenarios covering common questions and known failure cases, and a release that performs worse on those tests is not shipped.
Hallucinations can still occur. They are more likely when:
a document is very large or its text very small, particularly when the AI is asked to reproduce content word for word;
a conversation runs very long (see the limit below);
a single conversation involves many links and searches, which can lead to incorrect associations between papers or research areas.
Accuracy has not varied noticeably across research areas in testing. Research-area names are generated for each set of results and are not a fixed subject taxonomy.
Limits
Conversation length — a conversation can hold up to 200,000 tokens across all messages and uploaded content (a token is a unit of text; 200,000 is roughly 150,000 words). Past that, Chat prompts you to start a new conversation; shorter conversations also tend to be more accurate.
File uploads — PDFs up to 3 MB on the free tier and 20 MB on Keenious Plus. Uploaded files are deleted after one day.
These limits balance answer quality against the cost of running longer conversations, and may change over time.