Skip to content

v0.6.90b435

Choose a tag to compare

@github-actions github-actions released this 08 Sep 04:35
· 22 commits to main since this release
5212d4b

lilbee 0.6.90b435 clears the scanned-PDF known issue from b434, gives ingest an extraction timeout you can set, and repairs the confirm dialog.

Scanned PDFs read correctly again

b434 shipped with a known issue: on a scanned page, a numbered list came back as a three-column table and the page text appeared twice in the document. xberg 1.1.2 rejects a table candidate whose first column is list markers, and stops a detected table from repeating its text as paragraphs. The plain-output workaround lilbee carried for this is gone. On the scanned fixture, 1.1.2 returns the same text and no tables with or without it.

A per-file extraction timeout you can set

Every extraction ran under xberg's 600 second default, and nothing in lilbee could change it. LILBEE_EXTRACTION_TIMEOUT now sets that cap in seconds, on both extraction modes. Zero is the default and means no cap, which matches ocr_timeout and tesseract_timeout. A file that fails to extract also reports its reason now, instead of an object address.

The confirm dialog works again

The reset modal drew a blank box where "Yes" should be, and tabbing to "No" highlighted nothing. Yes and no are focusable pills again: the focused one takes the accent colour, and the arrow keys walk the row.

Two more model architectures

Architecture What it runs Upstream
hy_v4 Tencent Hy4-preview ggml-org/llama.cpp#28127
spark2_5 Spark-X2.5 ggml-org/llama.cpp#27868

lilbee now runs 150 GGUF architectures, up from 148. Pull any GGUF built on one of them and it works.
The engine changes are in dcaaea8d7...888307b81.

A release that finishes on its own

A publish leg that flakes is retried once instead of leaving a channel unpublished, and a dispatch that never landed is re-issued. Release shell moved out of the workflow files into scripts/, where shellcheck and a bats suite cover it.

Build and CI

The Ollama warm-up no longer segfaults on the GCC 11 AMX bug, and the catalog viewport tests wait for the grid to mount instead of flaking on Windows.

Full Changelog: v0.6.90b434...v0.6.90b435