[QinXi-ai]

Problem and behavior

Search could merge an original with a same-title repost. PDF imports could put license prose in the title, lose multiline titles and omit publication years. Chat retrieval omitted nearby explanations, and Web formulas appeared as plain Markdown.

Add independent arXiv discovery and prevent title-only/conflicting-DOI merges. PDF imports now combine geometry, optional source-located workspace-model extraction (including configured Qwen 27B), and exact arXiv/DOI lookup with title agreement. Conflicts and unknown fields remain explicit. Import UI/API reuse provider ownership, quota and dispatch protection; model failure has a visible fallback. Reimporting identical content can publish corrected metadata without downgrading an already verified version.

Chat context includes neighboring chunks/pages within a bounded source budget, preserves formula line breaks and renders KaTeX. Mixed valid/invalid citation answers are rejected, and location checks remain distinct from independent claim support. Download timeouts retain their type for existing pre-dispatch retries.

See docs/PDF_METADATA.md for configuration and boundaries. Existing libraries are corrected by reimport, not automatically rewritten.

Validation

• Backend: 234 passed, 1 existing skip; final targeted compatibility/regression checks recorded locally.
• Frontend: 68 tests, typecheck and build passed; lint has 0 errors and 3 existing Hook warnings.
• Real Attention latest/v1, RoFormer and FlashAttention-2 PDFs: full titles/authors and registry years 2017/2017/2021/2023. RoFormer's 2023 revision date did not replace its 2021 publication year.
• Playwright + Edge: real Attention/RoFormer upload, metadata revision and display passed with 0 page errors. Qwen selection/request wiring and worker dispatch/fallback used explicitly controlled model fixtures, not real inference.
• Real-PDF context checks cover omitted gradient/scaling explanations and throughput figures. A controlled answer verified KaTeX rendering and citation navigation.

Remaining limits

Real Qwen inference and real-model answer correctness are NOT_TESTED; no production credentials were used. Public arXiv PDF downloads remain intermittent (an actual read timeout was captured); upload succeeds and retry boundaries have automated coverage. Scanned PDF OCR and full formula-layout recovery are outside this change. Model source location is not semantic verification. No deployment or merge has been performed.