Give the model full retrieved chunks instead of 400-character snippets #83

2026-10-13T03:41:06-04:00

ryan commented

2026-10-13 03:41:06 -04:00

Context compaction replaces a tool's text with its artifact, and the search.documents artifact only carries 400-character UI snippets. So the full ~1,000-character chunk the retriever found never reached the model, and answers were built on fragments.

simba_search and obsidian_search_notes attach the full chunk text as model_context, the mechanism focused note reads already use. Compaction prefers it, shortens content under budget pressure, and keeps source ids. When no row fits, the ids survive in results_compaction.omitted_source_refs.
The persisted artifact and the UI are unchanged: persistence only stores data.
The per-chunk Source: {raw metadata dict} line (about a fifth of search output tokens) becomes Title | Date | Source. The Citation: /sources/<id> line is unchanged.

Review: Sonnet worker plus two review rounds. Round 1 fixed a failing tight-budget test and an import that made rag depend on conversation at module load. Round 2 checked the zero-row case; I added a test for it.

547 unit tests, Ruff.

🤖 Generated with Claude Code

Context compaction replaces a tool's text with its artifact, and the `search.documents` artifact only carries 400-character UI snippets. So the full ~1,000-character chunk the retriever found never reached the model, and answers were built on fragments. - `simba_search` and `obsidian_search_notes` attach the full chunk text as `model_context`, the mechanism focused note reads already use. Compaction prefers it, shortens `content` under budget pressure, and keeps source ids. When no row fits, the ids survive in `results_compaction.omitted_source_refs`. - The persisted artifact and the UI are unchanged: persistence only stores `data`. - The per-chunk `Source: {raw metadata dict}` line (about a fifth of search output tokens) becomes `Title | Date | Source`. The `Citation: /sources/<id>` line is unchanged. **Review:** Sonnet worker plus two review rounds. Round 1 fixed a failing tight-budget test and an import that made `rag` depend on `conversation` at module load. Round 2 checked the zero-row case; I added a test for it. 547 unit tests, Ruff. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

ryan added 1 commit 2026-10-13 03:41:07 -04:00

Give the model full retrieved chunks instead of 400-character snippets

CI / backend (pull_request) Successful in 41s

Details

CI / frontend (pull_request) Failing after 7s

Details

CI / backend (push) Successful in 42s

Details

CI / frontend (push) Failing after 6s

Details

b8ccc88b58

Compaction replaces a tool's text with its artifact, and the search artifact
only carries UI snippets, so the full chunk the retriever found never reached
the model. Search tools now attach the full text as model_context, which
compaction prefers and shortens under budget pressure while keeping source
ids; the persisted artifact and UI are unchanged. The per-chunk header drops
the raw metadata dict (about a fifth of search output tokens) for title, date
and source.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

ryan merged commit 8c0cedb3db into main

2026-10-13 03:41:12 -04:00

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: ryan/simbarag#83