docs(specs): table-aware parse-routing lever (PROPOSED, resource-deferred) - #12
Open
Zenetusken wants to merge 1 commit into
Open
docs(specs): table-aware parse-routing lever (PROPOSED, resource-deferred)#12Zenetusken wants to merge 1 commit into
Zenetusken wants to merge 1 commit into
Conversation
…rred) Spec for the lever the census-stats corpus surfaced + validated (PR #11): route table-dense born-digital docs to Docling/TableFormer instead of PyMuPDF (which mangles dense tables into run-on mega-cells), gated by a cheap find_tables density signal so Docling's ~12.6 min/doc cost is paid only when warranted. Empirics-first gate PASSED (census-stats: Docling tables → cell 8/10, aggregate 4/5, refusal_cf=1.0; PyMuPDF tables would false-refuse). HONEST on the key risk: this is a MUTATIVE re-render (Docling chunks != PyMuPDF chunks → re-ingest churns → ANN shift), the SAME eval-gated regression class as the form-linearizer (Increment C, measured-negative) and the query-aware truncation lever (measured-negative, reverted) — NOT "HARD-gate-neutral by construction" (correcting my PR#11 overclaim). The table-RAG win is proven on the dense docs; the suite impact is unproven → full-eval-gated. BUILD + the eval gate are RESOURCE-DEFERRED (the dev rig's VMs left no host-RAM headroom for the regression/validation sweep). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Spec for the parse-routing lever the census-stats corpus surfaced + validated (PR #11): route table-dense born-digital docs to Docling/TableFormer instead of PyMuPDF (which mangles dense tables), gated by a cheap
find_tablesdensity signal (Docling is ~12.6 min/doc, so density-gate it).Empirics PASSED (census-stats): Docling tables → cell 8/10, aggregate 4/5, refusal_cf=1.0; PyMuPDF tables would false-refuse.
Honest on the risk: this is a mutative re-render (re-ingest re-chunks → ANN shift) — the same eval-gated regression class as the form-linearizer + the just-reverted truncation lever, not HARD-gate-neutral by construction (correcting my PR#11 overclaim). The win is proven on the dense docs; the suite impact is unproven → full-eval-gated.
BUILD + eval are resource-deferred (host-RAM-blocked by the dev rig's VMs). Docs-only.
🤖 Generated with Claude Code