dsh-evidence

Getting a model in DeepSeek Harness to actually work with uploaded files has always meant one of two compromises: dump the entire document into the prompt and blow past the context window, or shell out to a Python script for every parse. dsh-evidence adds a third path. Files are parsed and indexed on your machine; the model retrieves only the compact evidence relevant to the question at hand, then expands the exact page, line range, or cell range it needs rather than pasting whole documents into the conversation.

It covers PDF text layers, DOCX body and tables, XLSX workbooks with A1 cell ranges, PPTX slides and speaker notes, common text encodings, and JPEG/PNG/WebP/GIF images via Harness' native vision path. CJK retrieval uses overlapping bigram indexing so Chinese phrases match correctly, and every coordinate is versioned with a source hash—edit the file and the stale coordinate fails loudly instead of silently returning the wrong lines. All parsing and indexing stay local; the plugin makes no external parsing calls.

If you regularly ask the model in the Harness Web composer to reason over contracts, research reports, spreadsheets, or slide decks without manually screenshotting or copy-pasting, this plugin is built for you. Upload once, ask a concrete question, and the model calls its own search and read tools on your behalf—saying the evidence was not found when it genuinely is absent, rather than guessing.