Camilla AI

How the discussion works
The user can work with a single AI model, any combination of the models, or all of them at once, and so adapt the analysis to the complexity of the question. The discussion runs for as many rounds as the content requires. Follow-up questions can build on earlier answers: the user can quote one or more answers in the next question, for example to ask the models to compare them, and every model has access to the conversation so far. An overview shows how much the models agree after each round, and a minimap along the conversation makes it easy to find your way back in long discussions. That several models agree is no proof that the answer is correct, because the models may share similar weaknesses. That is why the sources are shown, and the answers should always be checked against the originals.

Your own material
Camilla AI can combine the reasoning of the models with the user's own reference material through Retrieval-Augmented Generation (RAG). Scientific articles and other documents (PDF, Word, Excel and text files) are placed in a library, and Camilla AI retrieves the passages that are relevant to the question and gives them to the models, together with relevant earlier conversations. Earlier questions and answers are saved encrypted on the user's computer, so that new questions can build on what has already been discussed. The API keys of the models are also saved encrypted on the computer. A context panel shows exactly what was sent with each question.

Open databases and current information
When a question mentions a gene, a variant or a drug, Camilla AI looks it up in open databases (UniProt, ClinVar, gnomAD, Reactome, HPO, ChEMBL, AlphaFold, PDB, Ensembl and Orphanet, as well as drug labels from openFDA with DailyMed as a fallback, and NIH RePORTER when the question concerns funding) and gives the models the entries with identifiers and links, so that the answers can refer to real database entries. Web pages mentioned in a question are read directly and given to the models. For questions about information that can change over time, the models can also search the internet for current information before drawing conclusions.

Documents can be marked as private. They are then used only when the question concerns them, and the models' web search and database lookups are switched off, so that nothing from the document ends up in a search query. Files can also be marked as local only, so that they are never sent.

Literature monitoring
Literature monitoring lists newly published articles from PubMed, preprints from Europe PMC (which also serves as a fallback if PubMed does not respond) and new or updated clinical trials from ClinicalTrials.gov, straight from the sources and without AI interpretation. The searches are built automatically from the research area, or the user writes their own. The latest publications can also be fetched in advance, saved locally and given to the models in the discussions, so that the answers can build on current literature.

Autonomous discussions
If the user wishes, the models can discuss a chosen research area on their own, without the user having to ask every question. The user switches the mode on, and the setting is kept between sessions. The discussion starts when the user has been inactive for a set time and ends automatically after 30 minutes by default (30 or 60 minutes can be chosen). The models formulate new questions themselves and review each other's contributions through the usual discussion process, and the sources from the whole conversation are shown together at the end. The user can stop at any time and return to a normal conversation.

Voice, images and Excel
Questions can be asked by voice through a microphone. The spoken questions are written out in real time and shown for review before they are sent. Voice input currently supports English, and your own technical terms and abbreviations can be added so that the voice recognises them better. Answers from a chosen AI model can be read aloud, with speech and text synchronised and the language automatically matched to the user's question. A single answer can also be read aloud when needed, without links and source lists being read out.

Camilla AI can also create images and Excel files directly in the conversation: images are created by OpenAI, and spreadsheets are built by the AI model the user chooses. A conversation can also be saved as a PDF, with the colours of the models, the rounds, the formatting and the sources, and a single message or a whole conversation can be sent as an email to saved contacts through the user's own mail server. The interface is available in Swedish and English.

Models and statistics
The user chooses which model each provider should use and can give the models their own, shorter names in the window. A usage page shows how many questions and answers have been given over time, divided by model, and during the current session how fast each model answers and how often a call fails.

Areas of use
In medical research, Camilla AI can support discussions about disease mechanisms, genetics, metabolism, biological processes, published clinical observations, drugs and existing treatments. When several models review the same material, they can help identify connections, contradictions, knowledge gaps and possible new hypotheses. Hypotheses are suggestions for researchers to test, not results.

Important to know
Camilla AI is a research and analysis tool that complements human expertise. It is not intended to replace physicians, researchers or clinical judgement, and is not intended for diagnosis, treatment or decisions about individual patients. AI models can be wrong, and information and sources should always be checked against the originals.

Architecture