Checking installed local models...

Deep Reasoning automatically uses at least this many tokens of context. Increase if reasoning is still being cut off; decrease if you are running out of RAM. Qwen3 14B supports up to 131072.

Controls how the model's attention cache is stored in VRAM during generation. f16 (default) is full precision and uses the most memory. q8_0 cuts cache memory roughly in half with no meaningful quality difference — recommended if you are hitting VRAM limits or running long contexts. q4_0 halves it again but may slightly degrade coherence on very long answers. Requires Ollama 0.5 or newer; on older versions this setting is silently ignored.

Embedding models are cached on the VM. Building a new index preserves older indexes so you can switch back without losing the earlier ingestion.

Parent window size is fixed at 650 words (~850 tokens) with no overlap. Children are embedded; the matched child's parent window is returned to the LLM as context. Use Re-index clean after changing these settings.

Runtime: Tesseract, 1 thread per document. Worker status will appear when the VM reports telemetry.

Phase 2 runs on documents that Phase 1 (GLM-OCR) could not improve adequately — typically handwritten notes, degraded scans, tables, and mixed-language pages. Tesseract+ is the CPU fallback with layout-aware preprocessing; the VLM options use a vision language model to read the page image directly and produce much better results on difficult scans, but require a GPU with enough free VRAM to load an 8B model alongside the main LLM.

When enabled, every ingested PDF or image is checked against the low-yield threshold. Files that fall below it are automatically queued for vision OCR in the background. The pipeline pauses whenever you submit a query so the GPU stays responsive. Disable this if you want to control which files get re-processed manually.

A document is flagged as a candidate for vision OCR if its passage count divided by page count falls below this number. The default of 2 means a 10-page PDF with fewer than 20 indexed passages will be re-processed. Increase this to be more aggressive (re-process more files); decrease it to only catch near-blank pages.

Phase 1 (GLM-OCR) runs a fast vision pass on each page. Its result is only accepted — skipping the slower Phase 2 — if it produces at least this many times more words than the original extraction. 1.5 means GLM must be 50% better. Set lower (e.g. 1.1) to accept smaller gains and skip Phase 2 more often; set higher (e.g. 2.0) to demand clear improvement before trusting GLM's result.

A secondary acceptance criterion for Phase 1. Even if GLM only marginally improves on the original, if it produces a dense enough result (default: 80 words per page) it is assumed the page is well-covered and Phase 2 is skipped. Increase this if you are finding that GLM passes low-quality results; decrease it for shorter documents like cover pages or indexes.

CanLII metadata calls will be rate-limited and cached locally.