Spaces:

Hypn0
/

factory-rag

Configuration error

App Files Files Community

factory-rag

Commit History

Sep 11

09:30
fix: drop unused Path import (ruff F401) - project QC sweep; docs: sweep report

ML Assistant

c7a59dc
09:18
docs: porting guide additions - langextract, new flags, rebuild trigger, updated checklist

Hypn0

Verified signature
bcc25e0
06:26
docs: langextract audit + POC adoption + final Sept-2026 completeness statement

ML Assistant

16c8d3b
06:26
Add langextract field-extraction tab (optional, stub-tested)

ML Assistant

7eafc5d
06:14
docs: external-source challenge verdicts + adoptions

ML Assistant

2969255
06:14
feat: section_path metadata + section records, SME vocab expansion, optional pre-refusal gate (VDGR + doc-intel adoptions)

ML Assistant

6e49b19
03:59
docs: naive baseline decision entry

ML Assistant

916fce6
03:59
feat: naive fixed-chunk baseline strategy (eval + Compare panel 0) - the honest before picture

ML Assistant

3a0dab2

Sep 10

19:19
docs: final AB verdict - nemotron probes rank-0 4/4 verified, VL-reranker no-regression

ML Assistant

b909d1f
16:49
docs: Sept-2026 upgrade A/B measured results

ML Assistant

0c399ac
16:48
final AB driver: fixed probe doc-resolution + nemotron lane + VL-reranker before/after

Hypn0

Verified signature
dd829ae
15:16
fix: nemotron remote code needs datasets dep

ML Assistant

357a5c1
15:09
fix: drivers phase_fail unbound $2 crash on fail path

ML Assistant

ac0135c
15:08
test: CC-BY-NC variant A/B GPU driver (nemotron lane + jina listwise)

ML Assistant

09aa619
15:07
fix: nemotron build branch token-matrix bug (CPU roundtrip proof); eval k==5 double-count fix + MRR@5 metric; transformers>=4.57

ML Assistant

0dbf792
15:03
docs: license revision + NC variant decisions

ML Assistant

d4851c4
15:03
feat: NC best-in-class variants (nemotron-colembed lane, jina-reranker-v3.5 listwise); fix demo chips gitignored away

ML Assistant

38b7213
14:56
test: Sept-2026 upgrade A/B GPU driver

ML Assistant

a9946a9
14:54
feat(upgrades): zerank-2 text reranker tier + Tomoro ColQwen3 visual lane swap + Qwen3-VL-Reranker page reranking (A/B flags)

ML Assistant

c15ad20
14:41
docs: record doc summaries + quick ingest in decision log

ML Assistant

61bf81a
14:41
feat: per-document summaries (grader scope injection + summary records) and quick-ingest mode

ML Assistant

dd64696
12:55
docs: record demo chat upgrades in decision log

ML Assistant

ddc58a2
12:55
feat(demo): follow-up condensation, clickable demo chips, answer feedback logging

ML Assistant

ed6089f
10:14
Record real-TM validation campaign results (TM-5-811-6)

Hypn0

Verified signature
0472551
10:08
Add machine porting guide (Mac -> DGX Spark -> factory floor)

Hypn0

Verified signature
fc1b7e2
09:01
fix(enrich): dedicated VLM_NUM_CTX=8192 (16k KV alloc wedged daemon on image calls) + ollama.log dump on retries exhausted / breaker trip

ML Assistant

2667adb
08:32
fix(enrich): bounded VLM timeout (120s), page-call error containment, per-page progress logs, 3-failure circuit breaker -- one wedged daemon call must not hang or poison a large-document build

ML Assistant

aa19a3d
07:35
fix(smoke): disable tqdm/HF progress bars in job env (stalled log transport at embedding/visual stages)

ML Assistant

25be29a
04:24
feat(smoke): real-TM GPU driver (135-page US Army TM-5-811-6, vision lanes + LLM answer phases)

ML Assistant

e09b166
04:24
feat(eval): real set in EVAL_SETS (--set real, TM-5-811-6 manifest)

ML Assistant

6b37b37
04:12
feat(eval): real-document manifest from TM-5-811-6 (9 ground-truth rows, facts+pages verified in PDF)

ML Assistant

424e298
04:02
feat(app): side-by-side comparison panel (4 strategies, citations, page images, opt-in LLM answers); pin gradio 5.50.0

ML Assistant

92a3420
01:27
docs: hard-PDF validation campaign results (93.75% hard / 100% qa, GPU vision lanes 8/8, rotated-scan limitation)

ML Assistant

e4367cf
00:59
feat(scripts): hard-PDF GPU smoke driver

ML Assistant

4f455d8
00:52
feat(scripts): adversarial hard-PDF suite builder + CPU test overrides

ML Assistant

66b146f
00:52
feat(eval): --set hard with 16-row adversarial manifest (multi-column, tables, OCR scans, boilerplate)

ML Assistant

13e40ce
00:52
fix(ingest): thread per-doc doc_name through staged builds -- previously all chunks of a multi-doc ingest carried the last doc name (found by hard-PDF campaign)

ML Assistant

9e04784

Sep 9

20:35
docs: record post-validation bug audit in DECISIONS.md

ML Assistant

faf2d29
20:34
fix(ingest): page-range attribution per chunk (page+page_end) so cross-page chunks match page-exact queries; advisor detects externally deleted vector store

ML Assistant

971549e
20:26
chore(smoke): raise-from in driver; add stub test and synthetic PDF builder

ML Assistant

52e7dd0
20:26
chore(ingest_lib): drop dead converter branch and unused imports

ML Assistant

d8b5be6
20:26
chore: remove dead code in eval.py and retrieve.py (undefined-name line, unused prefetch)

ML Assistant

72cf4c7
20:26
fix(app): share chunks_holder global so page images attach; stream assistant bubble into history

ML Assistant

e1b1cad
19:06
Add validation-results section: CPU gates G1-G5 + GPU smoke v6 10/10 PASS

Hypn0

Verified signature
1da93d2
18:44
fix: pages collection as multivector (MaxSim, HNSW off) + deterministic page point IDs

Hypn0

579c183
18:23
fix: fpdf2 pin 2.8.8 (3.7.0 does not exist; stale sandbox copy overwrote the Hub fix)

Hypn0

72e4841
18:12
fix: visual_lane torch import, real HybridChunker usage, driver Phase B counts persisted vision records

Hypn0

97f7390
17:41
driver: install fpdf2, escape newlines in generated phase code, compile-check generated code

Hypn0

48caa8f
17:24
Add static landing page (Space hosts code; app runs locally)

Hypn0

Verified signature
20f58a4
17:24
eval: doc_name hit() + KX-40 eval rows + clean exit

Hypn0

8efaf2e

Commit History

Sep 11

fix: drop unused Path import (ruff F401) - project QC sweep; docs: sweep report

docs: porting guide additions - langextract, new flags, rebuild trigger, updated checklist

docs: langextract audit + POC adoption + final Sept-2026 completeness statement

Add langextract field-extraction tab (optional, stub-tested)

docs: external-source challenge verdicts + adoptions

feat: section_path metadata + section records, SME vocab expansion, optional pre-refusal gate (VDGR + doc-intel adoptions)

docs: naive baseline decision entry

feat: naive fixed-chunk baseline strategy (eval + Compare panel 0) - the honest before picture

Sep 10

docs: final AB verdict - nemotron probes rank-0 4/4 verified, VL-reranker no-regression

docs: Sept-2026 upgrade A/B measured results

final AB driver: fixed probe doc-resolution + nemotron lane + VL-reranker before/after

fix: nemotron remote code needs datasets dep

fix: drivers phase_fail unbound $2 crash on fail path

test: CC-BY-NC variant A/B GPU driver (nemotron lane + jina listwise)

fix: nemotron build branch token-matrix bug (CPU roundtrip proof); eval k==5 double-count fix + MRR@5 metric; transformers>=4.57

docs: license revision + NC variant decisions

feat: NC best-in-class variants (nemotron-colembed lane, jina-reranker-v3.5 listwise); fix demo chips gitignored away

test: Sept-2026 upgrade A/B GPU driver

feat(upgrades): zerank-2 text reranker tier + Tomoro ColQwen3 visual lane swap + Qwen3-VL-Reranker page reranking (A/B flags)

docs: record doc summaries + quick ingest in decision log

feat: per-document summaries (grader scope injection + summary records) and quick-ingest mode

docs: record demo chat upgrades in decision log

feat(demo): follow-up condensation, clickable demo chips, answer feedback logging

Record real-TM validation campaign results (TM-5-811-6)

Add machine porting guide (Mac -> DGX Spark -> factory floor)

fix(enrich): dedicated VLM_NUM_CTX=8192 (16k KV alloc wedged daemon on image calls) + ollama.log dump on retries exhausted / breaker trip

fix(enrich): bounded VLM timeout (120s), page-call error containment, per-page progress logs, 3-failure circuit breaker -- one wedged daemon call must not hang or poison a large-document build

fix(smoke): disable tqdm/HF progress bars in job env (stalled log transport at embedding/visual stages)

feat(smoke): real-TM GPU driver (135-page US Army TM-5-811-6, vision lanes + LLM answer phases)

feat(eval): real set in EVAL_SETS (--set real, TM-5-811-6 manifest)

feat(eval): real-document manifest from TM-5-811-6 (9 ground-truth rows, facts+pages verified in PDF)

feat(app): side-by-side comparison panel (4 strategies, citations, page images, opt-in LLM answers); pin gradio 5.50.0

docs: hard-PDF validation campaign results (93.75% hard / 100% qa, GPU vision lanes 8/8, rotated-scan limitation)

feat(scripts): hard-PDF GPU smoke driver

feat(scripts): adversarial hard-PDF suite builder + CPU test overrides

feat(eval): --set hard with 16-row adversarial manifest (multi-column, tables, OCR scans, boilerplate)

fix(ingest): thread per-doc doc_name through staged builds -- previously all chunks of a multi-doc ingest carried the last doc name (found by hard-PDF campaign)

Sep 9

docs: record post-validation bug audit in DECISIONS.md

fix(ingest): page-range attribution per chunk (page+page_end) so cross-page chunks match page-exact queries; advisor detects externally deleted vector store

chore(smoke): raise-from in driver; add stub test and synthetic PDF builder

chore(ingest_lib): drop dead converter branch and unused imports

chore: remove dead code in eval.py and retrieve.py (undefined-name line, unused prefetch)

fix(app): share chunks_holder global so page images attach; stream assistant bubble into history

Add validation-results section: CPU gates G1-G5 + GPU smoke v6 10/10 PASS

fix: pages collection as multivector (MaxSim, HNSW off) + deterministic page point IDs

fix: fpdf2 pin 2.8.8 (3.7.0 does not exist; stale sandbox copy overwrote the Hub fix)

fix: visual_lane torch import, real HybridChunker usage, driver Phase B counts persisted vision records

driver: install fpdf2, escape newlines in generated phase code, compile-check generated code

Add static landing page (Space hosts code; app runs locally)

eval: doc_name hit() + KX-40 eval rows + clean exit