Repository navigation

feat(perfs): add VLM-based picture/chart extraction to RAG eval pipeline - #57

Merged

ceberam merged 18 commits into

mainfrom

dev/rag-tree-opt

Sep 21, 2026

Commits on Sep 21, 2026

feat(rag): enable native heading hierarchy detection in conversion
ceberam
committed
b0d3f0f
fix(perfs): hierarchize and validate tree after conversion
ceberam
committed
a681e3d
feat(perfs): add picture_description option to Step 1 PDF conversion
ceberam
committed
8813634
refactor(perfs): drop step 2 (heading-level fix), renumber steps
ceberam
committed
0e7cab9
feat(perfs): add --chart-extraction option to Step 1
ceberam
committed
29033ed
feat(perfs): add --nemotron-ocr option to Step 1
ceberam
committed
b97464e
refactor(perfs): hard-code Nemotron OCR in Step 1
ceberam
committed
c0246a0
feat(perfs): extract _create_document_converter; infer Nemotron OCR language from dataset name
ceberam
committed
665e388
feat(perfs): save .dclx archive with page images alongside a lean JSON in Step 1
ceberam
committed
7230a0f
feat(perfs): per-file timing and status-aware logging in Step 1
ceberam
committed
81c75c0
feat(perfs): use Nemotron OCR on Linux, auto-select on other platforms
ceberam
committed
3fe5093
build: upgrade dependencies and add OCR dependencies
ceberam
committed
34469db
fix(rag): log parsing errors
ceberam
committed
85a1abb
fix: workaround on docling conversion with charts
ceberam
committed
9f00214
perf(eval): namespace step outputs by dataset name; disable classification
ceberam
committed
ebc4702
refactor(perfs): migrate chart extraction to ChartExtractionVlmEngineOptions with summary-only mode
ceberam
committed
4f56084
chore(perfs): customize picture description prompt to request a few sentences
ceberam
committed
db91809
feat(perfs): use Granite Vision 4.1-4B model spec for picture description
ceberam
committed
b0530f5

Navigation Menu

Repository navigation

feat(perfs): add VLM-based picture/chart extraction to RAG eval pipeline - #57

feat(perfs): add VLM-based picture/chart extraction to RAG eval pipeline#57
ceberam merged 18 commits into
mainfrom
dev/rag-tree-opt

Commits on Sep 21, 2026

feat(rag): enable native heading hierarchy detection in conversion

fix(perfs): hierarchize and validate tree after conversion

feat(perfs): add picture_description option to Step 1 PDF conversion

refactor(perfs): drop step 2 (heading-level fix), renumber steps

feat(perfs): add --chart-extraction option to Step 1

feat(perfs): add --nemotron-ocr option to Step 1

refactor(perfs): hard-code Nemotron OCR in Step 1

feat(perfs): extract _create_document_converter; infer Nemotron OCR language from dataset name

feat(perfs): save .dclx archive with page images alongside a lean JSON in Step 1

feat(perfs): per-file timing and status-aware logging in Step 1

feat(perfs): use Nemotron OCR on Linux, auto-select on other platforms

build: upgrade dependencies and add OCR dependencies

fix(rag): log parsing errors

fix: workaround on docling conversion with charts

perf(eval): namespace step outputs by dataset name; disable classification

refactor(perfs): migrate chart extraction to ChartExtractionVlmEngineOptions with summary-only mode

chore(perfs): customize picture description prompt to request a few sentences

feat(perfs): use Granite Vision 4.1-4B model spec for picture description