Encourages new topics. Penalizes any token that appeared regardless of count. 0 = disabled.
Max Tokens (num_predict)4096
Maximum number of tokens to generate per response. Higher = longer answers but slower. -1 = unlimited.
Context Length (num_ctx)32768
Context window size in tokens. Larger = more conversation memory but uses more VRAM.
Seed0
Random seed for reproducible outputs. 0 = random each time. Set a fixed number for consistent results.
Planning Max Attempts3
Maximum planning attempts before the agent proceeds. Higher = more retries if planning fails. 1-5 recommended.
Max Agent Steps25
Hard ceiling on reasoning steps per request. Each step is a full LLM call (~20s on a 24B model), so this caps worst-case wait time. Raise it only for genuinely multi-step research. 6–12 recommended.
Request Timeout (sec)300
Maximum seconds the agent may run before timing out. Increase for complex or long research tasks. 60–900 recommended.
📄 OCR (scans & images)
OCR EnabledOn
Pages with no text layer are rendered and sent to a vision model. Off means scanned PDFs extract nothing.
OCR Modeldeepseek-ocr:3b
Must be a vision model in Ollama. deepseek-ocr:3b is ~5x faster and more complete than qwen3-vl:8b on scans, and gets its own prompt format automatically.
Render DPI150
Resolution the PDF page is rasterised at before downscaling. Higher helps small print, costs time.
Max Image Edge1024
Longest edge sent to the model. Do NOT raise casually — qwen3-vl returns an EMPTY string well above 1024px, which looks like OCR failing silently.
Max Pages20
Safety cap on how many pages of one PDF get OCR'd.
OCR Timeout (s)180
Per-page limit on the vision model call.
Auto-Export Formatdocx
An OCR'd upload downloads as a document automatically. 'none' keeps it click-to-download from the Files panel.
🎤 Speech-to-Text (faster-whisper)
Whisper Modelsmall
Larger = more accurate, slower. 'small' is the sweet spot for short mic clips.
Devicecpu
cpu keeps speech off the GPU, which Ollama and the image pipelines already contend for.
Compute Typeint8
int8 on cpu, float16 on cuda. Changing the device resets this to match.
Voice Activity DetectionOn
Silero VAD strips silence before decoding. This is what stops Whisper inventing 'Thank you for watching.' on near-silent clips.
🎓 Tutor — Spaced Repetition
Schedulerfsrs
FSRS models difficulty, stability and retrievability per card and needs ~20-30% fewer reviews than SM-2 for the same retention. sm2 is the pre-2.6 fallback.
Desired Retention0.9
Target probability you still recall a card when it comes up. Lower = fewer, longer-spaced reviews.
Max Interval (days)365
Ceiling on any review gap. FSRS's own default is 36500 days, meaningless for language study.
Initial Ease (SM-2 only)2.5
Starting ease factor. Ignored when the scheduler is FSRS.
🎬 Video Generation
Video Backendwan
wan = Wan 2.2 TI2V-5B (default, best quality here — ~4.66B params, fits in VRAM at bf16 so there is no streaming penalty). ltx = LTX-Video 2B, faster but visibly weaker on anatomy under motion. Weights are 27-32GB and are never downloaded during a chat request. Switching backend also resets size/rate below to that model's native values, unless you pinned them in .env.
Width1280
Video is far more memory-hungry than images: every frame is generated jointly.
Height704
Keep modest on a 16GB card.
Frames Per Second24
LTX conditions on this: it tells the model how much movement belongs between frames. Generating at one rate and playing at another stretches the motion — 25 is LTX's native rate.
Duration (s)3
Clip length. Cost scales with duration.
Inference Steps50
More steps = better quality, linearly slower.
Guidance Scale5.0
How strictly the model follows the prompt.
Clips (stitched)1
Chain clips into one longer video — each continues from the LAST FRAME of the previous, so joins are continuous rather than cuts. Each clip adds ~2.7s. Cost is linear and TIME is the limit: ~7 min per clip at 720p, ~3.5 min at 832x480. A 1-minute video needs ~20 clips (about 70 min at 480p, 135 min at 720p). Quality also drifts over a long chain, since each clip is conditioned on the previous one. Wan only.
🔍 Image Self-Refinement
Refine Rounds2
After rendering, a vision model checks the image against your request and re-renders if it is wrong. Each round costs another generation plus a model swap, so images take roughly 2-3x longer. 0 = off (single shot, fastest).
Critic Modelqwen3-vl:8b
Must be a VISION model pulled in Ollama, and is independent of both the chat model and the image backend. If it is missing, refinement silently does nothing.