dsh-research

When using an AI agent for close reading of papers, a recurring pain point is that source quotes, the model own inferences, numeric normalizations, and cross-paper comparison conclusions blur into a single stream of output. After the fact, it is difficult to audit which sentence came from which physical page of which PDF, or to reliably distinguish source text from authored annotation. dsh-research embeds an evidence-first workflow into an existing DSH profile so that every citation carries physical-page, parser-revision, and quote-hash anchors, and so that inferences, comparison protocols, and synthesis conclusions are recorded as separate, traceable entities.

Core capabilities span three layers. First, local native-text PDF reading and retrieval: bounded excerpt collection over recognized section labels in English and Chinese, structural navigation via outlines, lexical search, and exact surrounding-block recovery. Second, a durable paper library that persists paper identities, source versions, parser observations, bibliography provenance, and reversible aliases in profile storage. Third, a cross-paper research-information workflow covering research questions, evidence, notes, claims, entities, observations, comparison protocols, synthesis matrices, and audit views. Synthesis inferences can retain explicit comparison protocols, and the renderer produces review-ready Markdown that clearly separates source summaries, inferences, comparison basis, and bibliography ledgers.

This bundle suits researchers working within a DSH profile with DeepSeek models who require rigorous citation and an auditable evidence chain, particularly those performing multi-paper comparisons, numeric result normalization, or structured literature reviews. The plugin does not perform semantic retrieval, OCR, automatic unit conversion, statistical significance testing, or meta-analysis. It provides structured tools so that the research process itself remains reproducible and auditable.