ankka-experiment-proposals

ankka-experiment-proposals is one ankka service that turns a customer request such as "Can you make foie gras?" into a reviewed experimental design. On the way it interprets the request, derives the scientific questions that must be answered, searches the literature and internal records, extracts findings and claims and checks each one against the text it came from, judges whether the evidence is enough for a stated decision, and proposes candidate experiments that a named scientist chooses between and approves.

It is a retrieval-augmented generation (RAG) application, and it is written as an example of doing RAG in a way that holds up: a language model is called at eleven points, and at every one of them its reply is a typed proposal that deterministic code checks before anything is recorded. Nothing a model says is taken as true because the model said it. The application can, and does, answer "insufficient evidence".

Where to start

You want to	Read
understand why it is built this way	Models inside a deterministic workflow
see how the parts fit together	How the application is built
see every stage, its check and its retry	The stages
understand the data types and who owns which decision	Proposals, records and rules
run it on your machine	Run it locally
follow a request end to end	The demonstration walkthrough
add a model-backed stage of your own	Add a model-backed stage
deploy it	Deploy on ankka

What it is not

It is a proof of concept for one research team, running in one installation with one instance. Its reviewers are names from a configured list rather than logins, its model replies in tests are recorded rather than live, and it parses JATS XML and plain text but not PDF. Limitations lists every such boundary.