An agentic RAG assistant (formerly "arXiv Assistant") that answers questions over the
RAG / agents / evaluation / LLMOps slice of arXiv, with citations. This glossary is the project's ubiquitous
language; decisions live in .scratch/arxiv-assistant/ and (eventually) SPEC.md.
Corpus: The set of papers matched by the pinned topical queries (RAG, LLM agents, LLM evaluation, LLMOps phrases; 2020→now, ~12.5k papers). Avoid: dataset, knowledge base
Full-text tier: The curated subset of the corpus (top-cited plus recent, ~2k papers) whose full text is fetched and indexed; all other papers are represented by abstract and metadata only.
Pinned snapshot: The corpus frozen at a recorded date — the committed paper-ID list and metadata. Evaluations and reviewer reproduction run against it.
Delta refresh: The scheduled re-run of the pinned queries that adds papers submitted since the last run to the live instance's corpus. Avoid: re-ingest, sync
Abstract card: The per-paper searchable unit: title, abstract, and key metadata as one text. Every corpus paper has exactly one.
Section chunk: A section-level passage of a full-text-tier paper, carrying its paper title and section heading. Only full-text-tier papers have them. Avoid: document, fragment
Evidence: The reranked set of abstract cards and section chunks an answer is grounded in and cites. Avoid: context, sources (in code and UI copy)
Evidence loop: The agent's cycle of formulating searches, gathering evidence, and synthesizing a cited answer — the demo's centerpiece. Avoid: pipeline, chain
Semantic search: The agent tool that retrieves evidence by meaning across abstract cards and section chunks (optionally scoped to one kind).
Metadata query: The agent tool that answers analytical questions (counts, filters, groupings, "what's new since X") over paper metadata via typed parameters. Avoid: structured query, SQL tool, text-to-SQL (that's the deferred v2 deepening)
Freshness query: A question about recent papers ("what's new in RAG eval this week?"), answered as a date-filtered metadata query plus synthesis.
Trace: The user-visible record of one evidence loop run: tool calls, evidence, and latency. Avoid: "How I answered" panel, log