What is field-proving-ground?

Field Proving Ground is a small local system for answering questions using retrieval augmented generation, a technique where a program looks up relevant passages from a fixed set of documents before answering, instead of relying purely on what an AI model already knows. The project's main goal is to make sure the system says it does not know when its source documents do not actually contain an answer, instead of quietly guessing based on the underlying model's own training. It is built to run entirely on modest, phone class hardware. It uses a small AI model called gemma2:2b for generating answers and a separate small model called nomic embed text for finding relevant passages in its document collection, referred to as the cold pack. The author is explicit that this is a genuinely constrained setup, not a larger model artificially limited to look small, and notes that a separate, bigger personal assistant project of theirs is unrelated to this one. The system has three main parts: a server script that only listens on the local machine and makes no outside network calls beyond talking to a locally running tool called Ollama, an evaluation script that runs a set of tests checking whether the system answers correctly or abstains appropriately, and the fixed folder of documents it draws its answers from. Before the model is allowed to answer, a separate check, called an abstain judge, looks at the retrieved passages and decides whether they truly contain the answer. If they do not, the system refuses to answer rather than filling the gap with its own guessed knowledge. The README documents real problems the author ran into while building this: early versions leaked outside knowledge because a simple similarity score let irrelevant passages through, and a single fixed similarity cutoff proved unreliable, which is why the abstain judge replaced it. To run it, you need Ollama installed locally with the two small models downloaded, plus a standard Python 3.9 or newer install with no extra packages required. You start the server with one command and run the test suite with another. The model is not kept loaded in memory between answers, and all traffic stays on your own machine.