Repository navigation
Replies: 2 comments
|
+1, and adding a second real-world case on a newer build (0.60.564.0, Postgres, 14 sources, personal brain). Our layout is the one this idea describes. All 29 person/company/project pages live in What we see today:
Duplicating people pages into every import source, or merging all the imports into The ask: the For context on why this matters to me: before GBrain I ran a plain LLM-wiki setup, and asking about a person pulled everything about them. GBrain is clearly better at growing and consolidating overnight, but losing the person-to-fact connection across imports is the biggest gap for us. Happy to run a preview build against our brain and report the numbers. |
|
Still reproducible on 0.60.545.0, recording the gap here as #6908 moved to this Discussion. Setup: a mail connector source (facts extracted from email pages) and a Measured on 0.60.545.0, free tiers only, dry run:
The 17 Oct dry run on 0.60.282.0 gave the same result (0 linked; 8,201 / 1,799), so nothing between those releases changed this. A paid re-extraction wouldn't help either, because subject inference stays source-scoped. Idea: mirror the loop design for facts. Add an optional subject source on the fact (e.g. |
Uh oh!
There was an error while loading. Please reload this page.
Moved from #6908, opened by @davidfragnoli-lang on 2026-10-17. Feature requests now live in Discussions > Ideas.
Version
0.60.282.0, Postgres, federated sources.
Observed
A fact can retain its evidence owner in source A (
facts.source_id,source_markdown_slug,row_num) but can identify its subject only with the bareentity_slug. All entity-scoped fact readers qualify that slug with the fact's owner source. If the canonical subject page exists only in federated source B, the fact cannot be attached to or retrieved from that entity without either duplicating the page into A or moving/copying the fact into B.gbrain facts relink --source A --no-llmis source-scoped and returnsno_pagefor these rows even when an exact, unique identity match (for example a unique email alias or anentity_identitiescanonical member) exists in B. The currentfactsschema has no explicit target source field, andgbrain repairhas no cross-source fact-target repair kind.This is the fact-store analogue of #4680, but a graph edge alone does not fix entity recall, facts fences, conflict sweeps, or provenance-preserving repair.
Expected
Support an explicit subject identity such as
(target_source_id, entity_slug)while preserving the fact's owner and evidence:facts.source_id,source_markdown_slug,row_num, visibility, trust, context, and fact id remain unchanged.Reproduction shape
entity_slugispeople/exampleand whose evidence remains in an A page.people/examplepage; federated source B has the canonical page and exact identity evidence.gbrain facts relink --source A --dry-run --no-llm --jsondoes not qualify B and reportsno_page/no link.No local patch is applied; this report follows the pure-upstream operating path.
All reactions