Evaluate hybrid graph and embedding retrieval for vault notes #1573

New issue

Closed

opened 2026-11-05 15:56:57 +00:00 by jared · 3 comments

jared commented

2026-11-05 15:56:57 +00:00

Owner

Problem

Jared asked whether a hybrid of graph search and embedding search would find vault notes better than graph search alone. No one has measured it. This ticket records the idea only. It has no design.

Jared's words, verbatim:

"Note search with embeddings is a better tool for this." This is interesting. I wonder how well it would perform if we:

queries the knowledge graph

embeddings query

unique returned notes list

rerank list

exclude low ranking matches (not sure how to best do this)

use context from remaining matches
These create extra steps, but we would be combining complimentary methods. And, if the result is lightning fast since we are using deterministic tooling, it could be worth it.

Need to dial-in our graphify first. Get it to where we are happy and confident in the results it is returning, before we consider adding other techniques.

This ticket waits for graphify sign-off on #1540. Do not start it before then. Capture applies no state labels, so the wait is stated here and no waiting label is set.

Reproduce

n/a: this is an idea to evaluate, and nothing is broken.

Origin

Trigger: Jared's remark during the graph retrieval work, session b10276a0.
Improvised this session: none.
Chain: vault retrieval needs a better method ← graph search alone returns weak results on reworded questions ← DESIGN (#1540).
Root candidate: none found; searched embedding, hybrid, graphify, vault note, retrieval
Where: plugins/os-vault
Session: b10276a0-3582-48b5-b0fe-f13795bdcc75
Transcript: /home/jared/.claude/projects/-home-jared-Documents-SecondBrain/b10276a0-3582-48b5-b0fe-f13795bdcc75.jsonl
Parent: #1540.

VERDICT: CREATE

## Problem Jared asked whether a hybrid of graph search and embedding search would find vault notes better than graph search alone. No one has measured it. This ticket records the idea only. It has no design. Jared's words, verbatim: > "Note search with embeddings is a better tool for this." This is interesting. I wonder how well it would perform if we: > - queries the knowledge graph > - embeddings query > - unique returned notes list > - rerank list > - exclude low ranking matches (not sure how to best do this) > - use context from remaining matches > These create extra steps, but we would be combining complimentary methods. And, if the result is lightning fast since we are using deterministic tooling, it could be worth it. > > Need to dial-in our graphify first. Get it to where we are happy and confident in the results it is returning, before we consider adding other techniques. This ticket waits for graphify sign-off on #1540. Do not start it before then. Capture applies no state labels, so the wait is stated here and no `waiting` label is set. ## Reproduce n/a: this is an idea to evaluate, and nothing is broken. ## Origin - Trigger: Jared's remark during the graph retrieval work, session b10276a0. - Improvised this session: none. - Chain: vault retrieval needs a better method ← graph search alone returns weak results on reworded questions ← DESIGN (#1540). - Root candidate: none found; searched embedding, hybrid, graphify, vault note, retrieval - Where: plugins/os-vault - Session: b10276a0-3582-48b5-b0fe-f13795bdcc75 - Transcript: /home/jared/.claude/projects/-home-jared-Documents-SecondBrain/b10276a0-3582-48b5-b0fe-f13795bdcc75.jsonl - Parent: #1540. VERDICT: CREATE

jared added the

plugin/os-vault

filed-by/user

category/investigation

labels

2026-11-05 15:56:57 +00:00

jared referenced this issue

2026-11-06 21:30:52 +00:00

Merge duplicate concept nodes and stop wikilink stubs taking the citing note as source #1568

jared commented

2026-11-06 22:07:58 +00:00

Author

Owner

Evaluated 2026-11-06 with a throwaway index: one card per note (title, file name, summary) in memsearch on bge-m3 CPU. 334 notes in 53 s. Expected note in top 5: original 11/11, reworded 12/13 (graph alone 6/13). Building it into vault-query on branch ticket-1573.

jared commented

2026-11-06 22:15:56 +00:00

Author

Owner

Resolution

Done: vault-query searches one card per note through memsearch, then adds graph notes. Disposable store in ~/.local/share/os-vault/cards, separate from the episodic store. No reranker: memsearch has none by default, and the top 5 already reach 12/13.

Evidence: main e65b7b68; results-after-1573: reworded 8/13 to 12/13, original 12/12, facts 18/25, no card failures. First run 75 s, later runs about 4 s. Rulings 72, 73.

Follow-ups: none

## Resolution **Done:** vault-query searches one card per note through memsearch, then adds graph notes. Disposable store in ~/.local/share/os-vault/cards, separate from the episodic store. No reranker: memsearch has none by default, and the top 5 already reach 12/13. **Evidence:** main e65b7b68; results-after-1573: reworded 8/13 to 12/13, original 12/12, facts 18/25, no card failures. First run 75 s, later runs about 4 s. Rulings 72, 73. **Follow-ups:** none

jared closed this issue

2026-11-06 22:15:57 +00:00

jared referenced this issue from a commit

2026-11-06 22:16:00 +00:00

docs(decisions): rule the #1573 card index design

jared referenced this issue from a commit

2026-11-06 22:16:00 +00:00

feat(os-vault): search note cards in vault-query (#1573)

jared referenced this issue from a commit

2026-11-06 22:17:59 +00:00

docs(os-vault): record the #1573 scores and set the summary rule (#1574)

jared commented