Adaptation Solutions Agentic Platform (ASAP), beta version: LLM agents for the transparent retrieval of quantitative adaptation costs from the literature

Abstract

Abstract. This paper describes the beta version of ASAP. Quantitative evidence on what climate adaptation measures cost is scattered across peer-reviewed studies, agency reports, national plans and, for small-scale measures, commercial price listings. Existing compilations are one-off, confined to a few sectors, and biased toward English-language and high-income sources, so adaptation enters risk assessments and integrated assessment models as an aggregate expenditure rather than as priced options. ASAP is a pipeline of stateless large-language-model (LLM) agents that retrieves documents from academic indexes, web search and commercial sources, screens them with a local embedding model and an LLM relevance gate, extracts every cost figure with a verbatim quote that is verified in code against the source text, harmonises figures to USD 2023 per canonical unit, and maps the free-text solution label onto a controlled vocabulary of 71 measures in 15 sectors. Every stored record carries its quote, page, source, reference year, currency and a computed confidence score, and every verdict that removes a document is reversible. This working paper documents the design and the evaluation protocol, describes the state of the first corpus, which so far covers only three of the fifteen vocabulary sectors (urban, energy demand and water) and is still being extracted, and sets out the form the final product takes: a traceable table of harmonised cost ranges per solution, country and sub-region with full source provenance. Measured against hand-built answer keys, the extractor finds about half of the printed cost figures with about the same precision, and the relevance gate cuts off-topic rows among the survivors by two thirds at a small loss of usable rows; these numbers are reported in full. The paper closes with the development steps that the first corpus has shown to be necessary, in particular a richer unit dictionary, capacity denominators for commercial listings and a finer level of the controlled vocabulary. Code and database are open.