Hacktoberfest 2026: những issue maintainer đã đánh dấu cho tháng Mười, đang mở và phù hợp người mới. Xem issue Hacktoberfest

epic: make the vector store a reconciled projection, then widen what enters it

Đang mở
  • Có gì không ổn?
#228 0 bình luận 0 reaction 1 người được giao Xem trên GitHub

Maintainer thường phản hồi trong vòng 1 ngày

@samarthsinh2660 đang làm issue này rồi.

Từ ngày 12/10/2026.

Đánh giá

Độ khó
5/5
Thời gian dự kiến
Hơn một tuần
Mức phù hợp với người mới
25/100
Loại issue
Tính năng
Độ rõ ràng
Đặc tả rõ ràng
Mức độ hoạt động
Sôi nổi
Công nghệ
dart
Lĩnh vực
databases

Hướng nghiên cứu

This is an umbrella issue and explicitly tracks no code of its own. Read docs/SHIVA/rag.md under “PDF documents” and the design at docs/superpowers/specs/2026-09-19-pdf-rag-design.md, then choose an open sub-issue (#231–#236) to work on. Done for this issue means its planned work is handled through those sub-issues; there is no standalone implementation or test target here.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.

Mô tả

enhancement P1-High

PDF content support landed in #226: a PDF attached to a note is now extracted, chunked, embedded and citable down to the page. Building it surfaced a set of gaps that are deliberately out of that issue's scope — some are pre-existing problems in the note side of the vector pipeline that only became visible once documents started retrieving reliably, others are capabilities the v1 design consciously deferred.

This issue is the umbrella for that follow-up work. It tracks no code of its own.

The idea

Retrieval is only as good as what made it into the vector store, and today the store is written once and never reconciled. A note embedded before the embedder existed stays unsearchable forever; a vector written by one tostore version can read back empty after a package bump; a chunk whose document was deleted leaves its vector behind with no way to reclaim it. None of these announce themselves — retrieval just quietly returns less.

The through-line for the next phases: make the vector store a reconciled, self-healing projection of what the user actually has, then widen what can enter it.

Ordered by what unblocks the most:

Phase A — make existing content searchable (correctness, the highest value work here)
Notes that never got a vector, and stores that silently read empty after an upgrade. Until this is fixed, every retrieval quality measurement is against an unknown fraction of the library.

Phase B — keep it healthy over time
Orphaned vectors, and the dimension the app declares versus the one the model emits.

Phase C — widen what documents contribute
A document's content reaches retrieval but not the knowledge graph, and a Manas-scoped chat cannot reach documents at all.

Phase D — reach and polish
Formats beyond PDF, and jumping to a cited page in-app.

Sub-issues

Phase A

  • #231 — Backfill: re-embed notes that have no vector
  • #232 — Version-stamp the vector store path so a package upgrade cannot read empty

Phase B

  • #233 — Rebuild path for orphaned chunk vectors
  • #234 — Reconcile the declared embedding dimension with what the model emits

Phase C

  • #235 — Documents contribute to the knowledge graph
  • #236 — A Manas-scoped chat should reach its notes' attached documents

Phase D

  • #237 — Jump to the cited page in-app
  • #238 — Support formats beyond PDF

Phase E — beyond text documents

  • #240 — Images: search the text inside attached images (on-device OCR)
  • #242 — Selective OCR for PDF and Word: scanned pages and pasted pictures only (measure on a device first)

Deliberately not planned

OCR for scanned documents. A PDF with no text layer stays attached, openable and marked not searchable. This is a product decision, not a backlog item — a scan's extracted text is unreliable enough that citing a page from it would be worse than not citing at all.

Reconsidered: OCR is now used for text inside images (#240), and is planned for scanned PDF pages and pictures pasted into Word files (#242) — page by page, never for a whole document, because a PDF's own text is exact and OCR's is not. #242 must be measured on a device before it is built.

Reference

  • Parent feature: #226
  • Design: docs/superpowers/specs/2026-09-19-pdf-rag-design.md
  • Current behaviour, with diagrams: docs/SHIVA/rag.md, section "PDF documents"
Ngôn ngữ chính
Dart
Star
2
Fork
2
Merge trung bình
18 giờ 53 phút
Pull request đã merge (30 ngày)
35

Chuẩn bị môi trường

Bắt đầu từ đâu

  1. Đọc hết issue, rồi đọc hướng dẫn đóng góp của dự án.
  2. Bình luận trên issue rằng bạn sẽ nhận — tránh hai người làm cùng một việc.
  3. Fork repository và làm thay đổi trên một nhánh.
  4. Mở pull request có tham chiếu số hiệu của issue.

Issue khác của basictech01/uniun

Tất cả issue của basictech01/uniun

Issue tương tự

Thêm issue về Dart

Nhận issue mới trong hộp thư của bạn

Bản tóm tắt ngắn những issue GitHub phù hợp với người mới.