EcoSearch is a retrieval-first system for transparent evidence access over prepared corpora. It combines semantic retrieval, lexical/anchor signals, reranking, confidence routing, and retrieval-quality evaluation to keep supporting evidence visible before any external generation layer is involved.
EcoSearch is designed around a simple principle: retrieve evidence first. Instead of starting from a chatbot response, the system searches a prepared corpus, surfaces supporting passages, and helps users inspect whether the retrieved text is strong enough to answer a question.
The project began as a retrieval-first QA workflow using chunking, embeddings, semantic search, reranking, confidence routing, and sentence-level evidence tracking. EcoSearch v1.1 extends this work with semantic/lexical balance experiments, rare-anchor evidence, oracle-overlap evaluation, and focus-hit diagnostics across multiple corpora.
The current mobile preview brings EcoSearch into a standalone app experience for iOS and Android. Prepared corpora can be queried directly inside the app, with retrieved evidence shown to the user. Corpus preparation can still use an OpenAI key, while the core retrieval workflow can run locally once the corpus has been prepared.
The goal is not to replace general-purpose chatbots. EcoSearch focuses on a narrower layer: retrieval quality, source visibility, answer containment, and grounded evidence access for document-based question answering.