Hey Product Hunt! I’m Aakash, the developer behind Trimwise. While building AI agents, I kept running into the same problem: large source documents, retrieved passages, search results, and tool outputs all had to fit alongside the actual instructions in a limited prompt budget. Like many developers, I started with simple prefix truncation: text[:N]. It was fast, but unreliable. The useful evidence might be near the beginning, buried in the middle, split across multiple sections, or just beyond N. I repeatedly ended up building custom chunking, ranking, and prompt-assembly logic for each project. Trimwise grew out of a simpler contract I wanted: Select the most relevant text from a known source, keep it verbatim, preserve its original order, and stay within an exact token, word, or character budget. It is an open-source Python library with: • lightweight lexical selection that requires no embedding model • optional semantic and hybrid selection • exact configurable budgets • source-backed excerpts instead of generated summaries • original character spans for tracing selected text back to its source It is designed as a prompt-assembly primitive for RAG systems, agents, document workflows, and other applications where provenance and predictable limits matter. I also released the benchmark cases, frozen outputs, scoring code, retrieval baselines, component studies, manuscript, and reproduction instructions. On the released 160-case author-annotated benchmark, the evaluated Trimwise configurations preserved complete required source spans more often than the specific compression adapters and fixed-window retrieval baselines tested. The paper also documents the failures and limitations rather than treating the benchmark as universal proof. I’d especially appreciate feedback on the API, real-world use cases, missing selection strategies, and cases where Trimwise chooses the wrong evidence. Thanks for checking it out!

Trimwise

Exact-budget, source-faithful context selection for prompts

Exact-budget, source-faithful context selection for prompts