Claude commited on
Commit
ec46cbf
·
1 Parent(s): 28435e9

Per-feature LLM toggles, section titles, redesigned upload page

Browse files

llm_extract: LLMToggles dataclass with alt_text/key_terms/math/section_title
flags. Disabled toggles drop the field from BOTH the prompt and the
Pydantic response schema (built dynamically via pydantic.create_model)
so we don't pay output tokens for fields we'd discard. Disabling all
of them skips the LLM call entirely. Section titles is a new field —
the LLM generates an ≤8-word descriptive title per segment, which the
orchestrator overrides on seg.section_title (replacing the generic
'Segment 3 — 4:25' default).

main.py: four new Form() bools (llm_alt_text/key_terms/math_equations/
section_title) plumbed into PipelineInputs.llm_toggles on both /jobs
and the legacy /process endpoint.

index.html: full redesign. Two-column hero with a polished form card +
side panel of features/tips. Drop-zone file inputs that show name,
size, video duration/resolution, transcript cue count + first cue
after selection. Output formats render as radio cards with
description copy. AI-assist section pairs the API key with four
toggle cards (one per LLM feature) so users can see what each costs
and turn off what they don't want. Tuning sliders moved into a
collapsed Advanced disclosure to reduce visual noise. Progress panel
gets a stage-pill row showing the pipeline phases as they advance.
Color palette + typography match the editor's maroon-and-serif look.

app/main.py CHANGED
@@ -17,6 +17,7 @@ from fastapi.templating import Jinja2Templates
17
 
18
  from .jobs import Job, JobResult, registry
19
  from .pipeline import bundle_zip, render_guide, render_review
 
20
  from .pipeline.orchestrator import (
21
  PipelineError,
22
  PipelineInputs,
@@ -175,6 +176,10 @@ async def start_job(
175
  subtitle: str | None = Form(None),
176
  module: str | None = Form(None),
177
  anthropic_api_key: str | None = Form(None),
 
 
 
 
178
  ) -> JSONResponse:
179
  video_suffix = Path(video.filename or "").suffix.lower()
180
  transcript_suffix = Path(transcript.filename or "").suffix.lower()
@@ -207,6 +212,12 @@ async def start_job(
207
  face_threshold=face_threshold,
208
  inline_images=(format in ("single", "review")),
209
  anthropic_api_key=(anthropic_api_key or None),
 
 
 
 
 
 
210
  )
211
 
212
  job = registry.create()
@@ -325,6 +336,10 @@ async def process(
325
  subtitle: str | None = Form(None),
326
  module: str | None = Form(None),
327
  anthropic_api_key: str | None = Form(None),
 
 
 
 
328
  ) -> Response:
329
  video_suffix = Path(video.filename or "").suffix.lower()
330
  transcript_suffix = Path(transcript.filename or "").suffix.lower()
@@ -358,6 +373,12 @@ async def process(
358
  face_threshold=face_threshold,
359
  inline_images=(format in ("single", "review")),
360
  anthropic_api_key=(anthropic_api_key or None),
 
 
 
 
 
 
361
  )
362
  try:
363
  result = run_pipeline(inputs)
 
17
 
18
  from .jobs import Job, JobResult, registry
19
  from .pipeline import bundle_zip, render_guide, render_review
20
+ from .pipeline.llm_extract import LLMToggles
21
  from .pipeline.orchestrator import (
22
  PipelineError,
23
  PipelineInputs,
 
176
  subtitle: str | None = Form(None),
177
  module: str | None = Form(None),
178
  anthropic_api_key: str | None = Form(None),
179
+ llm_alt_text: bool = Form(True),
180
+ llm_key_terms: bool = Form(True),
181
+ llm_math_equations: bool = Form(True),
182
+ llm_section_title: bool = Form(True),
183
  ) -> JSONResponse:
184
  video_suffix = Path(video.filename or "").suffix.lower()
185
  transcript_suffix = Path(transcript.filename or "").suffix.lower()
 
212
  face_threshold=face_threshold,
213
  inline_images=(format in ("single", "review")),
214
  anthropic_api_key=(anthropic_api_key or None),
215
+ llm_toggles=LLMToggles(
216
+ alt_text=llm_alt_text,
217
+ key_terms=llm_key_terms,
218
+ math_equations=llm_math_equations,
219
+ section_title=llm_section_title,
220
+ ),
221
  )
222
 
223
  job = registry.create()
 
336
  subtitle: str | None = Form(None),
337
  module: str | None = Form(None),
338
  anthropic_api_key: str | None = Form(None),
339
+ llm_alt_text: bool = Form(True),
340
+ llm_key_terms: bool = Form(True),
341
+ llm_math_equations: bool = Form(True),
342
+ llm_section_title: bool = Form(True),
343
  ) -> Response:
344
  video_suffix = Path(video.filename or "").suffix.lower()
345
  transcript_suffix = Path(transcript.filename or "").suffix.lower()
 
373
  face_threshold=face_threshold,
374
  inline_images=(format in ("single", "review")),
375
  anthropic_api_key=(anthropic_api_key or None),
376
+ llm_toggles=LLMToggles(
377
+ alt_text=llm_alt_text,
378
+ key_terms=llm_key_terms,
379
+ math_equations=llm_math_equations,
380
+ section_title=llm_section_title,
381
+ ),
382
  )
383
  try:
384
  result = run_pipeline(inputs)
app/pipeline/llm_extract.py CHANGED
@@ -18,7 +18,7 @@ from dataclasses import dataclass
18
  from pathlib import Path
19
  from typing import Callable
20
 
21
- from pydantic import BaseModel
22
 
23
  from .metadata import KeyTerm, MathEquation
24
 
@@ -28,36 +28,78 @@ DEFAULT_MODEL = "claude-opus-4-7"
28
  DEFAULT_MAX_CONCURRENCY = 4
29
  _API_KEY_ENV = "ANTHROPIC_API_KEY"
30
 
31
- # System prompt is byte-stable across all segments in a run, so prefix caching
32
- # kicks in after the first call. Keep dynamic content out of this string.
33
- _SYSTEM_PROMPT = """You analyze single frames from instructional video lectures and return structured JSON.
34
 
35
- For each frame you receive together with the spoken narration that accompanies it, extract:
 
 
 
 
 
 
 
 
 
36
 
37
- 1. alt_text — a WCAG-grade alt-text draft for the frame.
38
- - Convey the PURPOSE of the visual in the lecture's argument, not just a list of what's on screen.
39
- Example bad: "Slide containing the text 'Combining AI models, beyond monolithic AI'."
40
- Example good: "Title slide introducing the section on combining specialized AI models into systems
41
- that go beyond a single monolithic model."
42
- - One or two sentences, ≤200 characters. Plain prose, no markdown, no leading "Image of"/"Slide showing".
43
- - If the frame is purely an instructor talking-head with no slide content, say so plainly
44
- (e.g. "Instructor speaking to camera; no slide content on screen.").
45
- - Never invent details that aren't visible in the frame or supported by the narration.
46
 
47
- 2. key_terms — terms or named concepts that the lecturer is teaching in THIS segment.
48
- - Include only terms a student should remember or that are defined here.
49
- - Keep definitions concise (one sentence, ≤25 words). Plain prose, no markdown.
50
- - Skip generic English vocabulary, slide chrome, or section labels (e.g. "Outline", "Agenda").
51
- - Empty list if nothing on this slide rises to the bar.
52
 
53
- 3. math_equations — equations VISIBLY DISPLAYED on the slide.
54
- - latex: a clean LaTeX expression suitable for MathJax (no $ delimiters; no \\begin{equation}).
55
- - aria_label: a plain-English screen-reader description of the equation.
56
- - Skip equations that are only spoken, not on screen.
57
- - Skip simple inline labels like axis ticks or single numbers.
58
- - Empty list if the slide has no displayed equation.
59
 
60
- Return ONLY the structured fields. Do not invent content that isn't on the slide or in the narration."""
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
61
 
62
 
63
  class _ExtractedKeyTerm(BaseModel):
@@ -70,20 +112,37 @@ class _ExtractedMathEquation(BaseModel):
70
  aria_label: str
71
 
72
 
73
- class _SegmentExtraction(BaseModel):
74
- alt_text: str
75
- key_terms: list[_ExtractedKeyTerm]
76
- math_equations: list[_ExtractedMathEquation]
 
 
 
 
 
 
 
 
 
 
77
 
78
 
79
  @dataclass
80
  class ExtractionResult:
81
- alt_text: str
82
- key_terms: list[KeyTerm]
83
- math_equations: list[MathEquation]
 
 
 
 
 
 
 
84
 
85
 
86
- _EMPTY = ExtractionResult(alt_text="", key_terms=[], math_equations=[])
87
 
88
 
89
  def _resolve_api_key(api_key: str | None) -> str | None:
@@ -148,12 +207,16 @@ def extract_one(
148
  *,
149
  model: str = DEFAULT_MODEL,
150
  api_key: str | None = None,
 
151
  ) -> ExtractionResult:
152
- """Extract key terms + equations for a single segment. Empty on any failure.
153
 
154
- Returns an empty result (with a logged warning) on any error so a flaky
155
- network or schema-generation issue cannot break a video pipeline run.
 
156
  """
 
 
157
  resolved = _resolve_api_key(api_key)
158
  if resolved is None:
159
  return _EMPTY
@@ -165,7 +228,7 @@ def extract_one(
165
  system=[
166
  {
167
  "type": "text",
168
- "text": _SYSTEM_PROMPT,
169
  "cache_control": {"type": "ephemeral"},
170
  }
171
  ],
@@ -175,17 +238,22 @@ def extract_one(
175
  "content": _build_user_content(image_path, narration, ocr_text),
176
  }
177
  ],
178
- output_format=_SegmentExtraction,
179
  )
180
  parsed = response.parsed_output
181
  if parsed is None:
182
  log.warning("LLM extraction returned no parsed output for %s", image_path.name)
183
  return _EMPTY
184
  return ExtractionResult(
185
- alt_text=parsed.alt_text or "",
186
- key_terms=[KeyTerm(term=k.term, definition=k.definition) for k in parsed.key_terms],
 
 
 
 
187
  math_equations=[
188
- MathEquation(latex=m.latex, aria_label=m.aria_label) for m in parsed.math_equations
 
189
  ],
190
  )
191
  except Exception as exc: # pragma: no cover (network-dependent)
@@ -202,6 +270,7 @@ def extract_all(
202
  max_workers: int = DEFAULT_MAX_CONCURRENCY,
203
  on_progress: Callable[[int, int], None] | None = None,
204
  api_key: str | None = None,
 
205
  ) -> list[ExtractionResult]:
206
  """Run extract_one across all segments in parallel, preserving input order.
207
 
@@ -211,7 +280,7 @@ def extract_all(
211
  """
212
  n = len(image_paths)
213
  results: list[ExtractionResult] = [_EMPTY] * n
214
- if n == 0 or not is_enabled(api_key):
215
  if on_progress:
216
  on_progress(n, n)
217
  return results
@@ -227,6 +296,7 @@ def extract_all(
227
  ocr_texts[i] if i < len(ocr_texts) else "",
228
  model=model,
229
  api_key=api_key,
 
230
  ): i
231
  for i in range(n)
232
  }
 
18
  from pathlib import Path
19
  from typing import Callable
20
 
21
+ from pydantic import BaseModel, create_model
22
 
23
  from .metadata import KeyTerm, MathEquation
24
 
 
28
  DEFAULT_MAX_CONCURRENCY = 4
29
  _API_KEY_ENV = "ANTHROPIC_API_KEY"
30
 
 
 
 
31
 
32
+ @dataclass(frozen=True)
33
+ class LLMToggles:
34
+ """Per-feature opt-in flags. Disabling any one drops it from the LLM
35
+ request schema and prompt so the user is not billed for output tokens
36
+ they'd discard. Disabling all of them skips the LLM call entirely.
37
+ """
38
+ alt_text: bool = True
39
+ key_terms: bool = True
40
+ math_equations: bool = True
41
+ section_title: bool = True
42
 
43
+ @property
44
+ def any_enabled(self) -> bool:
45
+ return any((self.alt_text, self.key_terms, self.math_equations, self.section_title))
 
 
 
 
 
 
46
 
 
 
 
 
 
47
 
48
+ _DEFAULT_TOGGLES = LLMToggles()
 
 
 
 
 
49
 
50
+ # System prompt is byte-stable across all segments in a run, so prefix caching
51
+ # kicks in after the first call. Keep dynamic content out of this string.
52
+ _FIELD_PROMPTS = {
53
+ "section_title": (
54
+ "section_title — a short, descriptive title for this segment (≤8 words). "
55
+ "Capture the main idea, not the slide chrome. Sentence case, no trailing punctuation."
56
+ ),
57
+ "alt_text": (
58
+ "alt_text — a WCAG-grade alt-text draft for the frame.\n"
59
+ " - Convey the PURPOSE of the visual in the lecture's argument, not a transcription "
60
+ "of slide text. Bad: \"Slide containing 'Combining AI models'.\" "
61
+ "Good: \"Title slide introducing the section on combining specialized AI models.\"\n"
62
+ " - One or two sentences, ≤200 characters. Plain prose, no markdown, no leading "
63
+ "\"Image of\"/\"Slide showing\".\n"
64
+ " - If the frame is purely an instructor talking-head with no slide content, say so plainly.\n"
65
+ " - Never invent details that aren't visible in the frame or supported by the narration."
66
+ ),
67
+ "key_terms": (
68
+ "key_terms — terms or named concepts that the lecturer is teaching in THIS segment.\n"
69
+ " - Include only terms a student should remember or that are defined here.\n"
70
+ " - Keep definitions concise (one sentence, ≤25 words). Plain prose, no markdown.\n"
71
+ " - Skip generic English vocabulary or section labels (\"Outline\", \"Agenda\").\n"
72
+ " - Empty list if nothing on this slide rises to the bar."
73
+ ),
74
+ "math_equations": (
75
+ "math_equations — equations VISIBLY DISPLAYED on the slide.\n"
76
+ " - latex: a clean LaTeX expression suitable for MathJax (no $ delimiters; "
77
+ "no \\begin{equation}).\n"
78
+ " - aria_label: a plain-English screen-reader description of the equation.\n"
79
+ " - Skip equations that are only spoken, not on screen, or simple inline labels.\n"
80
+ " - Empty list if the slide has no displayed equation."
81
+ ),
82
+ }
83
+
84
+ _SYSTEM_PREAMBLE = (
85
+ "You analyze single frames from instructional video lectures and return structured JSON.\n\n"
86
+ "For each frame you receive together with the spoken narration that accompanies it, extract:\n\n"
87
+ )
88
+
89
+ _SYSTEM_POSTAMBLE = (
90
+ "\n\nReturn ONLY the structured fields. Do not invent content that isn't on the slide "
91
+ "or in the narration."
92
+ )
93
+
94
+
95
+ def _build_system_prompt(toggles: LLMToggles) -> str:
96
+ parts = []
97
+ n = 1
98
+ for key in ("section_title", "alt_text", "key_terms", "math_equations"):
99
+ if getattr(toggles, key):
100
+ parts.append(f"{n}. {_FIELD_PROMPTS[key]}")
101
+ n += 1
102
+ return _SYSTEM_PREAMBLE + "\n\n".join(parts) + _SYSTEM_POSTAMBLE
103
 
104
 
105
  class _ExtractedKeyTerm(BaseModel):
 
112
  aria_label: str
113
 
114
 
115
+ def _build_extraction_model(toggles: LLMToggles):
116
+ """Pydantic model whose fields match the enabled toggles, so the LLM
117
+ schema only requests fields we'll actually use.
118
+ """
119
+ fields: dict = {}
120
+ if toggles.section_title:
121
+ fields["section_title"] = (str, ...)
122
+ if toggles.alt_text:
123
+ fields["alt_text"] = (str, ...)
124
+ if toggles.key_terms:
125
+ fields["key_terms"] = (list[_ExtractedKeyTerm], ...)
126
+ if toggles.math_equations:
127
+ fields["math_equations"] = (list[_ExtractedMathEquation], ...)
128
+ return create_model("_SegmentExtraction", **fields)
129
 
130
 
131
  @dataclass
132
  class ExtractionResult:
133
+ alt_text: str = ""
134
+ key_terms: list[KeyTerm] = None # type: ignore[assignment]
135
+ math_equations: list[MathEquation] = None # type: ignore[assignment]
136
+ section_title: str = ""
137
+
138
+ def __post_init__(self) -> None:
139
+ if self.key_terms is None:
140
+ self.key_terms = []
141
+ if self.math_equations is None:
142
+ self.math_equations = []
143
 
144
 
145
+ _EMPTY = ExtractionResult()
146
 
147
 
148
  def _resolve_api_key(api_key: str | None) -> str | None:
 
207
  *,
208
  model: str = DEFAULT_MODEL,
209
  api_key: str | None = None,
210
+ toggles: LLMToggles = _DEFAULT_TOGGLES,
211
  ) -> ExtractionResult:
212
+ """Extract enabled fields for a single segment. Empty on any failure.
213
 
214
+ Disabled toggles are dropped from both the prompt and the response schema
215
+ so we don't pay for tokens we'd discard. Any per-segment failure logs a
216
+ warning and returns _EMPTY — never aborts the pipeline.
217
  """
218
+ if not toggles.any_enabled:
219
+ return _EMPTY
220
  resolved = _resolve_api_key(api_key)
221
  if resolved is None:
222
  return _EMPTY
 
228
  system=[
229
  {
230
  "type": "text",
231
+ "text": _build_system_prompt(toggles),
232
  "cache_control": {"type": "ephemeral"},
233
  }
234
  ],
 
238
  "content": _build_user_content(image_path, narration, ocr_text),
239
  }
240
  ],
241
+ output_format=_build_extraction_model(toggles),
242
  )
243
  parsed = response.parsed_output
244
  if parsed is None:
245
  log.warning("LLM extraction returned no parsed output for %s", image_path.name)
246
  return _EMPTY
247
  return ExtractionResult(
248
+ alt_text=getattr(parsed, "alt_text", "") or "",
249
+ section_title=getattr(parsed, "section_title", "") or "",
250
+ key_terms=[
251
+ KeyTerm(term=k.term, definition=k.definition)
252
+ for k in getattr(parsed, "key_terms", []) or []
253
+ ],
254
  math_equations=[
255
+ MathEquation(latex=m.latex, aria_label=m.aria_label)
256
+ for m in getattr(parsed, "math_equations", []) or []
257
  ],
258
  )
259
  except Exception as exc: # pragma: no cover (network-dependent)
 
270
  max_workers: int = DEFAULT_MAX_CONCURRENCY,
271
  on_progress: Callable[[int, int], None] | None = None,
272
  api_key: str | None = None,
273
+ toggles: LLMToggles = _DEFAULT_TOGGLES,
274
  ) -> list[ExtractionResult]:
275
  """Run extract_one across all segments in parallel, preserving input order.
276
 
 
280
  """
281
  n = len(image_paths)
282
  results: list[ExtractionResult] = [_EMPTY] * n
283
+ if n == 0 or not is_enabled(api_key) or not toggles.any_enabled:
284
  if on_progress:
285
  on_progress(n, n)
286
  return results
 
296
  ocr_texts[i] if i < len(ocr_texts) else "",
297
  model=model,
298
  api_key=api_key,
299
+ toggles=toggles,
300
  ): i
301
  for i in range(n)
302
  }
app/pipeline/orchestrator.py CHANGED
@@ -13,7 +13,7 @@ from pathlib import Path
13
  from typing import Callable
14
 
15
  from .filter import filter_instructor_frames
16
- from .llm_extract import extract_all, is_enabled as llm_enabled
17
  from .metadata import PageMetadata, SourceInfo, page_metadata_from_segments
18
  from .ocr import ocr_frame
19
  from .scenes import detect_and_extract
@@ -54,6 +54,7 @@ class PipelineInputs:
54
  inline_images: bool = False
55
  # Per-job credential for LLM extraction; never logged or persisted.
56
  anthropic_api_key: str | None = field(default=None, repr=False)
 
57
 
58
 
59
  @dataclass
@@ -143,14 +144,22 @@ def run_pipeline(
143
  pct = 65 + ((i + 1) / max(n, 1)) * 30
144
  _emit(pct, "ocr", f"OCR frame {i + 1}/{n}")
145
 
146
- if llm_enabled(inputs.anthropic_api_key):
147
  n = len(kept)
148
- log.info("Running LLM extraction on %d frames", n)
149
- _emit(80, "llm_extract", f"Extracting key terms and equations ({n} frames)...")
 
 
 
 
 
 
 
 
150
 
151
  def _llm_progress(done: int, total: int) -> None:
152
  pct = 80 + (done / max(total, 1)) * 15
153
- _emit(pct, "llm_extract", f"Key terms / equations {done}/{total}")
154
 
155
  extractions = extract_all(
156
  [f.image_path for f in kept],
@@ -158,6 +167,7 @@ def run_pipeline(
158
  ocr_texts,
159
  on_progress=_llm_progress,
160
  api_key=inputs.anthropic_api_key,
 
161
  )
162
  else:
163
  extractions = None
@@ -166,10 +176,14 @@ def run_pipeline(
166
  segments = build_segments(kept, narrations, ocr_texts, inline_images=inputs.inline_images)
167
  if extractions is not None:
168
  for seg, extracted in zip(segments, extractions):
169
- seg.key_terms = list(extracted.key_terms)
170
- seg.math_equations = list(extracted.math_equations)
171
- if extracted.alt_text:
 
 
172
  seg.alt_draft = extracted.alt_text
 
 
173
 
174
  page = page_metadata_from_segments(
175
  title=inputs.title,
 
13
  from typing import Callable
14
 
15
  from .filter import filter_instructor_frames
16
+ from .llm_extract import LLMToggles, extract_all, is_enabled as llm_enabled
17
  from .metadata import PageMetadata, SourceInfo, page_metadata_from_segments
18
  from .ocr import ocr_frame
19
  from .scenes import detect_and_extract
 
54
  inline_images: bool = False
55
  # Per-job credential for LLM extraction; never logged or persisted.
56
  anthropic_api_key: str | None = field(default=None, repr=False)
57
+ llm_toggles: LLMToggles = field(default_factory=LLMToggles)
58
 
59
 
60
  @dataclass
 
144
  pct = 65 + ((i + 1) / max(n, 1)) * 30
145
  _emit(pct, "ocr", f"OCR frame {i + 1}/{n}")
146
 
147
+ if llm_enabled(inputs.anthropic_api_key) and inputs.llm_toggles.any_enabled:
148
  n = len(kept)
149
+ enabled_names = [
150
+ name for name, on in [
151
+ ("titles", inputs.llm_toggles.section_title),
152
+ ("alt-text", inputs.llm_toggles.alt_text),
153
+ ("key terms", inputs.llm_toggles.key_terms),
154
+ ("equations", inputs.llm_toggles.math_equations),
155
+ ] if on
156
+ ]
157
+ log.info("Running LLM extraction on %d frames (%s)", n, ", ".join(enabled_names))
158
+ _emit(80, "llm_extract", f"Extracting {', '.join(enabled_names)} ({n} frames)...")
159
 
160
  def _llm_progress(done: int, total: int) -> None:
161
  pct = 80 + (done / max(total, 1)) * 15
162
+ _emit(pct, "llm_extract", f"LLM extraction {done}/{total}")
163
 
164
  extractions = extract_all(
165
  [f.image_path for f in kept],
 
167
  ocr_texts,
168
  on_progress=_llm_progress,
169
  api_key=inputs.anthropic_api_key,
170
+ toggles=inputs.llm_toggles,
171
  )
172
  else:
173
  extractions = None
 
176
  segments = build_segments(kept, narrations, ocr_texts, inline_images=inputs.inline_images)
177
  if extractions is not None:
178
  for seg, extracted in zip(segments, extractions):
179
+ if inputs.llm_toggles.key_terms:
180
+ seg.key_terms = list(extracted.key_terms)
181
+ if inputs.llm_toggles.math_equations:
182
+ seg.math_equations = list(extracted.math_equations)
183
+ if inputs.llm_toggles.alt_text and extracted.alt_text:
184
  seg.alt_draft = extracted.alt_text
185
+ if inputs.llm_toggles.section_title and extracted.section_title:
186
+ seg.section_title = extracted.section_title
187
 
188
  page = page_metadata_from_segments(
189
  title=inputs.title,
app/templates/index.html CHANGED
@@ -3,344 +3,718 @@
3
  <head>
4
  <meta charset="utf-8">
5
  <meta name="viewport" content="width=device-width, initial-scale=1">
6
- <title>VideoGuideMaker</title>
7
- <link rel="stylesheet" href="/static/style.css">
 
 
8
  <style>
9
- .vgm-progress {
10
- display: none;
11
- margin: 1.25rem 0;
12
- padding: 1rem 1.25rem;
13
- background: #F5F5F5;
14
- border: 1px solid #D0D0D0;
15
- border-radius: 6px;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16
  }
17
- .vgm-progress[data-state="running"], .vgm-progress[data-state="error"] { display: block; }
18
- .vgm-progress h2 { font-size: 1.05rem; margin: 0 0 0.5rem; }
19
- .vgm-progress progress {
20
- width: 100%;
21
- height: 1rem;
22
- accent-color: #750014;
 
 
23
  }
24
- .vgm-progress__msg {
25
- margin-top: 0.5rem;
26
- font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", monospace;
27
- font-size: 0.95rem;
28
- color: #333;
 
 
29
  }
30
- .vgm-progress__stage {
 
 
 
 
 
 
 
 
 
31
  display: inline-block;
32
  font-size: 0.75rem;
33
  font-weight: 700;
34
  text-transform: uppercase;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35
  letter-spacing: 0.08em;
36
- color: #fff;
37
- background: #750014;
38
- padding: 0.1rem 0.5rem;
39
- border-radius: 3px;
40
- margin-right: 0.4rem;
41
- vertical-align: middle;
42
  }
43
- .vgm-progress[data-state="error"] {
44
- background: #fde2e2;
45
- border-color: #f5b5b5;
 
46
  }
47
- .vgm-progress[data-state="error"] .vgm-progress__stage { background: #7a0a0a; }
48
- .vgm-progress__details {
49
- display: none;
50
- margin-top: 0.5rem;
51
- padding: 0.5rem 0.75rem;
52
- background: #fff;
53
- border: 1px solid #f5b5b5;
54
- border-radius: 4px;
55
- font-family: ui-monospace, "SFMono-Regular", Menlo, Consolas, monospace;
56
- font-size: 0.8rem;
57
- color: #5a0808;
58
- white-space: pre-wrap;
59
- word-break: break-word;
60
- max-height: 14rem;
61
- overflow: auto;
 
62
  }
63
- .vgm-progress[data-state="error"] .vgm-progress__details:not(:empty) { display: block; }
64
- .vgm-progress__retry {
65
- display: none;
66
- margin-top: 0.6rem;
67
- padding: 0.4rem 0.9rem;
68
- background: #750014;
69
- color: #fff;
70
- border: 0;
71
- border-radius: 4px;
72
- font-size: 0.9rem;
 
 
 
 
 
73
  cursor: pointer;
 
 
74
  }
75
- .vgm-progress[data-state="error"] .vgm-progress__retry { display: inline-block; }
76
- .vgm-review {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77
  display: none;
78
- position: fixed;
79
- inset: 0;
80
- background: #fff;
81
- z-index: 100;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
82
  flex-direction: column;
 
 
83
  }
84
- .vgm-review[data-state="open"] { display: flex; }
85
- .vgm-review__bar {
86
- flex: 0 0 auto;
 
 
 
 
 
 
 
 
 
 
 
 
 
87
  display: flex;
88
- align-items: center;
89
- gap: 0.75rem;
90
- padding: 0.5rem 1rem;
91
- background: #1a1a1a;
92
- color: #fff;
93
- border-bottom: 1px solid #333;
94
- font-size: 0.9rem;
 
95
  }
96
- .vgm-review__bar strong { font-weight: 600; }
97
- .vgm-review__bar button {
98
- background: transparent;
99
- border: 1px solid #555;
 
 
 
 
 
 
 
 
 
 
 
 
 
100
  color: #fff;
101
- padding: 0.3rem 0.7rem;
102
- border-radius: 4px;
103
- font-size: 0.85rem;
 
 
104
  cursor: pointer;
 
 
105
  }
106
- .vgm-review__bar button:hover { background: #2a2a2a; }
107
- .vgm-review__bar .spacer { flex: 1; }
108
- .vgm-review__frame {
109
- flex: 1 1 auto;
110
- width: 100%;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
111
  border: 0;
 
112
  }
113
- .vgm-apikey-help {
114
- display: block;
115
- margin-top: 0.25rem;
116
- font-size: 0.8rem;
117
- color: #555;
 
 
 
 
 
118
  }
 
 
 
 
 
 
 
 
 
 
 
 
119
  .vgm-tune {
120
- margin: 0.9rem 0 1.1rem;
121
- padding-bottom: 0.6rem;
122
- border-bottom: 1px solid #eee;
123
  }
124
  .vgm-tune:last-of-type { border-bottom: 0; }
125
  .vgm-tune__head {
126
- display: flex;
127
- align-items: baseline;
128
- justify-content: space-between;
129
- gap: 0.6rem;
130
- margin-bottom: 0.25rem;
131
- }
132
- .vgm-tune__head label {
133
- font-weight: 600;
134
- font-size: 0.95rem;
135
  }
 
136
  .vgm-tune__value {
137
- font-family: ui-monospace, "SFMono-Regular", Menlo, Consolas, monospace;
138
- font-size: 0.9rem;
139
- color: #333;
140
- background: #f3f3f3;
141
  padding: 0.1rem 0.5rem;
142
  border-radius: 4px;
143
  min-width: 4ch;
144
  text-align: center;
145
  }
146
- .vgm-tune input[type="range"] {
147
- width: 100%;
148
- accent-color: #750014;
149
- margin: 0.15rem 0;
150
- }
151
  .vgm-tune input[type="range"]:disabled { opacity: 0.45; }
152
  .vgm-tune__anchors {
153
- display: flex;
154
- justify-content: space-between;
155
- font-size: 0.72rem;
156
- color: #777;
157
- letter-spacing: 0.02em;
158
- text-transform: uppercase;
159
  margin-top: -0.1rem;
160
  }
161
  .vgm-tune__help {
162
- display: block;
163
- margin-top: 0.35rem;
164
- font-size: 0.82rem;
165
- color: #555;
166
- line-height: 1.4;
167
  }
168
  .vgm-tune__toggle {
169
- display: inline-flex;
170
- align-items: center;
171
- gap: 0.4rem;
172
- font-size: 0.85rem;
173
- margin-bottom: 0.25rem;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
174
  }
 
 
 
175
  </style>
176
  </head>
177
  <body>
178
  <a class="skip-link" href="#main">Skip to main content</a>
179
- <header role="banner">
180
- <h1>VideoGuideMaker</h1>
181
- <p>Generate WCAG-ready HTML study guides from instructional videos.</p>
182
- </header>
183
 
184
- <main id="main" tabindex="-1">
185
- {% if error %}
186
- <div class="error" role="alert">{{ error }}</div>
187
- {% endif %}
 
 
 
 
 
 
 
 
188
 
189
  <div class="vgm-progress" data-state="idle" role="status" aria-live="polite" aria-atomic="true">
190
- <h2>Generating study guide…</h2>
191
  <progress max="100" value="0"></progress>
192
  <p class="vgm-progress__msg">
193
  <span class="vgm-progress__stage">starting</span>
194
  <span class="vgm-progress__text">Uploading…</span>
195
  </p>
 
 
 
 
 
 
 
 
 
196
  <pre class="vgm-progress__details" aria-label="Error details"></pre>
197
  <button type="button" class="vgm-progress__retry">Try again</button>
198
  </div>
199
 
200
- <form id="vgm-form" action="/process" method="post" enctype="multipart/form-data" aria-labelledby="form-heading">
201
- <h2 id="form-heading">Upload video and transcript</h2>
202
 
203
- <div class="field">
204
- <label for="title">Study guide title</label>
205
- <input type="text" id="title" name="title" value="Lecture Study Guide" required>
206
- </div>
207
 
208
- <div class="field">
209
- <label for="video">Video file (.mp4, .mov, .mkv, .webm)</label>
210
- <input type="file" id="video" name="video" accept="video/*" required>
211
- </div>
212
 
213
- <div class="field">
214
- <label for="transcript">Transcript (.vtt or .srt)</label>
215
- <input type="file" id="transcript" name="transcript" accept=".vtt,.srt,text/vtt,application/x-subrip" required>
216
- </div>
217
 
218
- <fieldset>
219
- <legend>Output format</legend>
220
- <label><input type="radio" name="format" value="review" checked> Review &amp; edit in browser (recommended)</label>
221
- <label><input type="radio" name="format" value="zip"> Zip bundle (HTML + static folder)</label>
222
- <label><input type="radio" name="format" value="single"> Single self-contained HTML (images inlined)</label>
223
- </fieldset>
224
-
225
- <fieldset>
226
- <legend>Key terms &amp; equations (optional)</legend>
227
- <label for="anthropic_api_key">Anthropic API key</label>
228
- <input type="password" id="anthropic_api_key" name="anthropic_api_key"
229
- autocomplete="off" spellcheck="false"
230
- placeholder="sk-ant-..."
231
- aria-describedby="anthropic_api_key_help">
232
- <small id="anthropic_api_key_help" class="vgm-apikey-help">
233
- Used once for this run to extract key terms and on-screen equations as LaTeX.
234
- Sent to Anthropic over HTTPS, held in memory only, never written to disk or logs.
235
- Leave blank to skip — the rest of the pipeline still runs.
236
- </small>
237
- </fieldset>
238
-
239
- <fieldset>
240
- <legend>Tuning</legend>
241
-
242
- <div class="vgm-tune">
243
- <div class="vgm-tune__head">
244
- <label for="threshold">Scene-change sensitivity</label>
245
- <output for="threshold" class="vgm-tune__value">27</output>
246
  </div>
247
- <input type="range" id="threshold" name="threshold" value="27" min="5" max="60" step="1">
248
- <div class="vgm-tune__anchors"><span>More scenes</span><span>Fewer scenes</span></div>
249
- <small class="vgm-tune__help">
250
- How visually different two frames must be before they count as a new scene.
251
- Lower (5–15) catches every animation step; higher (35+) only major slide changes.
252
- Default 27 works for most lecture decks.
253
- </small>
254
- </div>
255
 
256
- <div class="vgm-tune">
257
- <div class="vgm-tune__head">
258
- <label for="min_gap">Minimum gap between scenes</label>
259
- <output for="min_gap" class="vgm-tune__value">off</output>
 
 
 
 
 
 
260
  </div>
261
- <input type="range" id="min_gap" name="min_gap" value="0" min="0" max="60" step="0.5">
262
- <div class="vgm-tune__anchors"><span>Off</span><span>60 s</span></div>
263
- <small class="vgm-tune__help">
264
- Drop new scenes that arrive less than this many seconds after the previous kept one.
265
- Useful when a slide builds up gradually — set to ~5 s to keep just the final state.
266
- </small>
267
  </div>
268
 
269
- <div class="vgm-tune">
270
- <div class="vgm-tune__head">
271
- <label for="max_frames">Max frames</label>
272
- <output for="max_frames" class="vgm-tune__value">unlimited</output>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
273
  </div>
274
- <label class="vgm-tune__toggle">
275
- <input type="checkbox" id="max_frames_enabled"> Cap total frames
276
- </label>
277
- <input type="range" id="max_frames" name="max_frames" min="1" max="200" step="1" value="50" disabled>
278
- <div class="vgm-tune__anchors"><span>1</span><span>200</span></div>
279
- <small class="vgm-tune__help">
280
- Hard limit on the number of frames in the guide. Useful for very long lectures or
281
- to bound LLM extraction cost. Frames are evenly distributed across the video.
282
- </small>
283
  </div>
284
 
285
- <div class="vgm-tune">
286
- <div class="vgm-tune__head">
287
- <label for="face_threshold">Instructor-frame face threshold</label>
288
- <output for="face_threshold" class="vgm-tune__value">0.12</output>
 
 
 
 
 
 
 
 
 
 
 
289
  </div>
290
- <input type="range" id="face_threshold" name="face_threshold" value="0.12" min="0.02" max="0.6" step="0.01">
291
- <div class="vgm-tune__anchors"><span>Strict (drop more)</span><span>Lenient (keep more)</span></div>
292
- <small class="vgm-tune__help">
293
- Fraction of the frame a detected face must occupy before we treat it as
294
- instructor-only and drop it. Lower values discard talking-head shots more
295
- aggressively; raise it if your slides have a small inset webcam you want to keep.
296
- </small>
297
  </div>
298
 
299
- <div class="vgm-tune">
300
- <label class="vgm-tune__toggle">
301
- <input type="checkbox" name="skip_ocr" value="true"> Skip OCR
302
- </label>
303
- <small class="vgm-tune__help">
304
- Faster; on-screen text and the heuristic LaTeX wrapping in the guide will be empty.
305
- Doesn't affect LLM extraction of equations from the frame image.
306
- </small>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
307
  </div>
308
 
309
- <div class="vgm-tune">
310
- <div class="vgm-tune__head">
311
- <label for="lang">Document language</label>
312
- <output for="lang" class="vgm-tune__value">en</output>
313
- </div>
314
- <input type="text" id="lang" name="lang" value="en" pattern="[a-zA-Z\-]+" maxlength="12"
315
- style="width: 6rem;">
316
- <small class="vgm-tune__help">
317
- BCP-47 code (e.g. <code>en</code>, <code>en-US</code>, <code>fr</code>).
318
- Sets the <code>lang</code> attribute on the generated HTML and selects Tesseract
319
- language packs when available.
320
- </small>
321
  </div>
322
- </fieldset>
323
-
324
- <button type="submit">Generate study guide</button>
325
- </form>
326
-
327
- <section aria-labelledby="how-heading">
328
- <h2 id="how-heading">How it works</h2>
329
- <ol>
330
- <li>Detects scene changes in your video.</li>
331
- <li>Extracts a representative frame for each scene.</li>
332
- <li>Filters out talking-head frames using face detection.</li>
333
- <li>Runs OCR on each retained frame.</li>
334
- <li>Aligns transcript narration to each visual segment.</li>
335
- <li>Produces a WCAG 2.2 AA-ready HTML document with draft alt-text flagged for human review.</li>
336
- </ol>
337
- <p>
338
- In <strong>review mode</strong>, the result opens in your browser as an editor: rewrite alt text
339
- and section titles, edit narration and on-screen text, drop unwanted segments, then download a
340
- clean final HTML. Edits can be saved and reloaded as JSON.
341
- </p>
342
- </section>
343
- </main>
 
 
 
 
 
344
 
345
  <div class="vgm-review" data-state="closed" role="dialog" aria-label="Study guide editor">
346
  <div class="vgm-review__bar">
@@ -362,6 +736,7 @@
362
  const textEl = panel.querySelector('.vgm-progress__text');
363
  const detailsEl = panel.querySelector('.vgm-progress__details');
364
  const retryBtn = panel.querySelector('.vgm-progress__retry');
 
365
  const apiKeyEl = form.querySelector('#anthropic_api_key');
366
  const titleEl = form.querySelector('#title');
367
  const reviewEl = document.querySelector('.vgm-review');
@@ -371,6 +746,16 @@
371
  const reviewOpenBtn = reviewEl.querySelector('.vgm-review__open');
372
  let lastResultUrl = null;
373
 
 
 
 
 
 
 
 
 
 
 
374
  function openReview(resultUrl, title) {
375
  lastResultUrl = resultUrl;
376
  reviewTitleEl.textContent = title ? '— ' + title : '';
@@ -388,7 +773,89 @@
388
  if (lastResultUrl) window.open(lastResultUrl, '_blank', 'noopener');
389
  });
390
 
391
- // Tuning panel: keep each slider's <output> in sync with its value.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
392
  function fmtMinGap(v) { return parseFloat(v) <= 0 ? 'off' : (parseFloat(v) + ' s'); }
393
  function fmtIdentity(v) { return v; }
394
  const formatters = { min_gap: fmtMinGap };
@@ -400,7 +867,6 @@
400
  slider.addEventListener('input', sync);
401
  sync();
402
  });
403
- // Plain text input ('lang') gets the same readout treatment.
404
  form.querySelectorAll('input[type="text"][id]').forEach(input => {
405
  const out = form.querySelector('output[for="' + input.id + '"]');
406
  if (!out) return;
@@ -409,8 +875,7 @@
409
  sync();
410
  });
411
 
412
- // Max-frames toggle: when unchecked the slider is disabled and its value
413
- // is omitted from the form so the server treats it as "unlimited".
414
  const maxFramesToggle = form.querySelector('#max_frames_enabled');
415
  const maxFramesSlider = form.querySelector('#max_frames');
416
  const maxFramesOut = form.querySelector('output[for="max_frames"]');
@@ -434,21 +899,21 @@
434
  function setProgress(state, percent, stage, message) {
435
  panel.setAttribute('data-state', state);
436
  if (typeof percent === 'number') bar.value = percent;
437
- if (stage) stageEl.textContent = stage.replace('_', ' ');
 
 
 
438
  if (message) textEl.textContent = message;
439
  }
440
-
441
  function showError(short, detail) {
442
  setProgress('error', 100, 'error', short);
443
  detailsEl.textContent = detail || '';
444
  }
445
-
446
  function resetForRetry() {
447
  detailsEl.textContent = '';
448
  panel.setAttribute('data-state', 'idle');
449
  form.style.display = '';
450
  }
451
-
452
  retryBtn.addEventListener('click', resetForRetry);
453
 
454
  function postJob(fd) {
@@ -477,8 +942,10 @@
477
  const v = fd.get(k);
478
  if (typeof v === 'string' && v.trim() === '') fd.delete(k);
479
  });
480
- // Clear the API key from the DOM as soon as it's in the FormData; the
481
- // FormData copy is sent over HTTPS and discarded with this closure.
 
 
482
  if (apiKeyEl) apiKeyEl.value = '';
483
  const submittedTitle = (titleEl && titleEl.value) || 'Study Guide';
484
  const formatRadio = form.querySelector('input[name="format"]:checked');
@@ -500,7 +967,7 @@
500
  try {
501
  const parsed = JSON.parse(detail);
502
  detail = parsed.detail || parsed.message || detail;
503
- } catch (_) { /* keep raw text */ }
504
  showError(`Server rejected the upload (HTTP ${xhr.status})`, detail);
505
  return;
506
  }
@@ -524,10 +991,9 @@
524
  setProgress('running', data.percent, data.stage, data.message);
525
  } else if (data.type === 'done') {
526
  es.close();
527
- setProgress('running', 100, 'done', 'Loading editor…');
528
  if (format === 'review') {
529
  openReview(result_url, submittedTitle);
530
- // Reset the form pane so it's ready when the user clicks "Back to upload".
531
  panel.setAttribute('data-state', 'idle');
532
  form.style.display = '';
533
  } else {
@@ -536,7 +1002,7 @@
536
  document.body.appendChild(a);
537
  a.click();
538
  a.remove();
539
- setProgress('running', 100, 'done', 'Downloaded. Generate another?');
540
  setTimeout(resetForRetry, 2500);
541
  }
542
  } else if (data.type === 'error') {
@@ -545,8 +1011,6 @@
545
  }
546
  };
547
  es.onerror = () => {
548
- // EventSource will retry automatically on transient drops; only treat
549
- // as fatal when the underlying connection is closed permanently.
550
  if (es.readyState === EventSource.CLOSED) {
551
  showError('Lost connection to server.',
552
  `Last progress message: ${lastStageMsg}\n` +
 
3
  <head>
4
  <meta charset="utf-8">
5
  <meta name="viewport" content="width=device-width, initial-scale=1">
6
+ <title>Video Guide Maker</title>
7
+ <link rel="preconnect" href="https://fonts.googleapis.com">
8
+ <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
9
+ <link href="https://fonts.googleapis.com/css2?family=Fraunces:wght@500;600;700&family=Inter:wght@400;500;600;700&display=swap" rel="stylesheet">
10
  <style>
11
+ :root {
12
+ --color-primary: #750014;
13
+ --color-primary-dark: #5a000f;
14
+ --color-primary-soft: #f0e3e5;
15
+ --color-text: #1a1a1a;
16
+ --color-text-muted: #555;
17
+ --color-text-subtle: #888;
18
+ --color-border: #d8d8dc;
19
+ --color-border-soft: #ececef;
20
+ --color-bg: #fafaf7;
21
+ --color-surface: #ffffff;
22
+ --color-surface-alt: #f5f1ec;
23
+ --color-error-bg: #fde2e2;
24
+ --color-error-fg: #7a0a0a;
25
+ --space-xs: 0.4rem;
26
+ --space-sm: 0.6rem;
27
+ --space-md: 1rem;
28
+ --space-lg: 1.5rem;
29
+ --space-xl: 2.4rem;
30
+ --shadow-sm: 0 1px 2px rgba(0,0,0,0.04), 0 1px 3px rgba(0,0,0,0.05);
31
+ --shadow-md: 0 2px 6px rgba(0,0,0,0.05), 0 8px 24px rgba(0,0,0,0.06);
32
+ --radius-sm: 6px;
33
+ --radius-md: 10px;
34
+ --font-heading: 'Fraunces', Georgia, 'Times New Roman', serif;
35
+ --font-body: 'Inter', -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif;
36
+ --font-mono: ui-monospace, "SFMono-Regular", Menlo, Consolas, monospace;
37
  }
38
+ * { box-sizing: border-box; }
39
+ body {
40
+ margin: 0;
41
+ background: var(--color-bg);
42
+ color: var(--color-text);
43
+ font-family: var(--font-body);
44
+ line-height: 1.55;
45
+ font-size: 17px;
46
  }
47
+ h1, h2, h3, h4 { font-family: var(--font-heading); letter-spacing: -0.01em; line-height: 1.2; }
48
+ a { color: var(--color-primary); }
49
+ :focus-visible { outline: 2px solid var(--color-primary); outline-offset: 2px; border-radius: 4px; }
50
+
51
+ .skip-link {
52
+ position: absolute; left: -9999px; top: 0;
53
+ background: var(--color-primary); color: #fff; padding: 0.5rem 1rem; z-index: 1000;
54
  }
55
+ .skip-link:focus { left: 1rem; top: 1rem; }
56
+
57
+ .vgm-page {
58
+ max-width: 72rem;
59
+ margin: 0 auto;
60
+ padding: 2.5rem 1.5rem;
61
+ }
62
+
63
+ /* HEADER */
64
+ .vgm-eyebrow {
65
  display: inline-block;
66
  font-size: 0.75rem;
67
  font-weight: 700;
68
  text-transform: uppercase;
69
+ letter-spacing: 0.12em;
70
+ color: var(--color-primary);
71
+ margin-bottom: 0.75rem;
72
+ }
73
+ h1.vgm-title {
74
+ font-size: clamp(2.1rem, 4vw, 3rem);
75
+ font-weight: 600;
76
+ margin: 0 0 0.5rem;
77
+ }
78
+ h1.vgm-title em { font-style: italic; color: var(--color-primary); font-weight: 500; }
79
+ .vgm-tagline {
80
+ font-size: 1.1rem;
81
+ color: var(--color-text-muted);
82
+ margin: 0 0 var(--space-xl);
83
+ max-width: 40rem;
84
+ line-height: 1.5;
85
+ }
86
+
87
+ /* TWO-COLUMN HERO */
88
+ .vgm-hero {
89
+ display: grid;
90
+ grid-template-columns: minmax(0, 1.6fr) minmax(18rem, 1fr);
91
+ gap: 2rem;
92
+ align-items: start;
93
+ }
94
+ @media (max-width: 920px) {
95
+ .vgm-hero { grid-template-columns: 1fr; }
96
+ }
97
+
98
+ /* CARD */
99
+ .vgm-card {
100
+ background: var(--color-surface);
101
+ border: 1px solid var(--color-border);
102
+ border-radius: var(--radius-md);
103
+ padding: var(--space-lg);
104
+ box-shadow: var(--shadow-sm);
105
+ }
106
+ .vgm-card + .vgm-card { margin-top: 1rem; }
107
+
108
+ .vgm-section { padding: var(--space-md) 0; }
109
+ .vgm-section + .vgm-section { border-top: 1px solid var(--color-border-soft); }
110
+ .vgm-section:first-child { padding-top: 0; }
111
+ .vgm-section:last-child { padding-bottom: 0; }
112
+ .vgm-section__title {
113
+ font-size: 0.78rem;
114
+ font-weight: 700;
115
+ text-transform: uppercase;
116
  letter-spacing: 0.08em;
117
+ color: var(--color-text-muted);
118
+ margin: 0 0 var(--space-sm);
119
+ font-family: var(--font-body);
 
 
 
120
  }
121
+ .vgm-section__intro {
122
+ margin: -0.2rem 0 0.75rem;
123
+ color: var(--color-text-muted);
124
+ font-size: 0.92rem;
125
  }
126
+
127
+ /* INPUTS */
128
+ .vgm-field { display: flex; flex-direction: column; gap: 0.3rem; margin-bottom: var(--space-md); }
129
+ .vgm-field:last-child { margin-bottom: 0; }
130
+ .vgm-field label { font-weight: 600; font-size: 0.92rem; }
131
+ .vgm-field small { font-size: 0.82rem; color: var(--color-text-muted); }
132
+ .vgm-card input[type="text"],
133
+ .vgm-card input[type="password"] {
134
+ font-family: inherit;
135
+ font-size: 1rem;
136
+ padding: 0.6rem 0.75rem;
137
+ border: 1px solid var(--color-border);
138
+ border-radius: var(--radius-sm);
139
+ background: var(--color-surface);
140
+ width: 100%;
141
+ transition: border-color 0.15s, box-shadow 0.15s;
142
  }
143
+ .vgm-card input[type="text"]:focus,
144
+ .vgm-card input[type="password"]:focus {
145
+ outline: none;
146
+ border-color: var(--color-primary);
147
+ box-shadow: 0 0 0 3px var(--color-primary-soft);
148
+ }
149
+
150
+ /* DROP ZONES */
151
+ .vgm-drop {
152
+ position: relative;
153
+ display: block;
154
+ padding: 1.4rem 1rem;
155
+ border: 2px dashed var(--color-border);
156
+ border-radius: var(--radius-md);
157
+ text-align: center;
158
  cursor: pointer;
159
+ transition: border-color 0.15s, background 0.15s;
160
+ background: var(--color-bg);
161
  }
162
+ .vgm-drop:hover, .vgm-drop:focus-within {
163
+ border-color: var(--color-primary);
164
+ background: var(--color-surface-alt);
165
+ }
166
+ .vgm-drop.is-dragover {
167
+ border-color: var(--color-primary);
168
+ background: var(--color-primary-soft);
169
+ }
170
+ .vgm-drop input[type="file"] {
171
+ position: absolute; inset: 0;
172
+ width: 100%; height: 100%;
173
+ opacity: 0; cursor: pointer;
174
+ }
175
+ .vgm-drop__hint { font-size: 0.95rem; color: var(--color-text-muted); }
176
+ .vgm-drop__hint strong { color: var(--color-text); font-weight: 600; }
177
+ .vgm-drop__hint small { display: block; font-size: 0.8rem; color: var(--color-text-subtle); margin-top: 0.2rem; }
178
+ .vgm-drop__meta {
179
  display: none;
180
+ font-family: var(--font-mono);
181
+ font-size: 0.85rem;
182
+ text-align: left;
183
+ }
184
+ .vgm-drop__meta strong { color: var(--color-primary); font-weight: 600; word-break: break-all; }
185
+ .vgm-drop__meta dl {
186
+ display: grid;
187
+ grid-template-columns: max-content 1fr;
188
+ gap: 0.1rem 0.5rem;
189
+ margin: 0.5rem 0 0;
190
+ font-size: 0.78rem;
191
+ color: var(--color-text-muted);
192
+ }
193
+ .vgm-drop__meta dt { font-weight: 600; }
194
+ .vgm-drop[data-loaded="true"] {
195
+ border-style: solid;
196
+ border-color: var(--color-primary);
197
+ background: var(--color-surface);
198
+ }
199
+ .vgm-drop[data-loaded="true"] .vgm-drop__hint { display: none; }
200
+ .vgm-drop[data-loaded="true"] .vgm-drop__meta { display: block; }
201
+
202
+ /* OUTPUT FORMAT */
203
+ .vgm-formats {
204
+ display: grid;
205
+ grid-template-columns: repeat(auto-fit, minmax(11rem, 1fr));
206
+ gap: 0.5rem;
207
+ }
208
+ .vgm-fmt {
209
+ cursor: pointer;
210
+ padding: 0.7rem 0.85rem;
211
+ border: 1px solid var(--color-border);
212
+ border-radius: var(--radius-sm);
213
+ transition: border-color 0.15s, background 0.15s;
214
+ display: flex;
215
  flex-direction: column;
216
+ gap: 0.2rem;
217
+ background: var(--color-surface);
218
  }
219
+ .vgm-fmt:hover { border-color: var(--color-primary); }
220
+ .vgm-fmt:has(input:checked) {
221
+ border-color: var(--color-primary);
222
+ background: var(--color-primary-soft);
223
+ }
224
+ .vgm-fmt__head { display: flex; align-items: center; gap: 0.5rem; font-weight: 600; font-size: 0.93rem; }
225
+ .vgm-fmt small { font-size: 0.78rem; color: var(--color-text-muted); }
226
+
227
+ /* LLM TOGGLE GRID */
228
+ .vgm-llm-grid {
229
+ display: grid;
230
+ grid-template-columns: repeat(auto-fit, minmax(12rem, 1fr));
231
+ gap: 0.5rem;
232
+ margin-top: var(--space-sm);
233
+ }
234
+ .vgm-llm-toggle {
235
  display: flex;
236
+ gap: 0.55rem;
237
+ align-items: flex-start;
238
+ padding: 0.6rem 0.7rem;
239
+ border: 1px solid var(--color-border);
240
+ border-radius: var(--radius-sm);
241
+ cursor: pointer;
242
+ background: var(--color-surface);
243
+ transition: border-color 0.15s, background 0.15s;
244
  }
245
+ .vgm-llm-toggle:hover { border-color: var(--color-primary); }
246
+ .vgm-llm-toggle:has(input:checked) {
247
+ border-color: var(--color-primary);
248
+ background: var(--color-primary-soft);
249
+ }
250
+ .vgm-llm-toggle input { margin-top: 0.2rem; flex: 0 0 auto; accent-color: var(--color-primary); }
251
+ .vgm-llm-toggle__body { display: flex; flex-direction: column; gap: 0.1rem; min-width: 0; }
252
+ .vgm-llm-toggle strong { font-size: 0.9rem; font-weight: 600; }
253
+ .vgm-llm-toggle small { font-size: 0.75rem; color: var(--color-text-muted); line-height: 1.3; }
254
+
255
+ /* SUBMIT */
256
+ .vgm-cta {
257
+ display: inline-flex;
258
+ align-items: center;
259
+ gap: 0.5rem;
260
+ padding: 0.85rem 1.7rem;
261
+ background: var(--color-primary);
262
  color: #fff;
263
+ font-family: inherit;
264
+ font-weight: 600;
265
+ font-size: 1.02rem;
266
+ border: 0;
267
+ border-radius: var(--radius-sm);
268
  cursor: pointer;
269
+ box-shadow: var(--shadow-sm);
270
+ transition: background 0.15s, transform 0.05s;
271
  }
272
+ .vgm-cta:hover { background: var(--color-primary-dark); }
273
+ .vgm-cta:active { transform: translateY(1px); }
274
+ .vgm-submit-row { display: flex; justify-content: flex-end; align-items: center; gap: 1rem; margin-top: var(--space-md); }
275
+ .vgm-submit-row__hint { font-size: 0.82rem; color: var(--color-text-muted); }
276
+
277
+ /* SIDE PANEL */
278
+ .vgm-side h3 { font-size: 1.05rem; margin: 0 0 0.7rem; }
279
+ .vgm-feature-list { list-style: none; padding: 0; margin: 0; display: flex; flex-direction: column; gap: 0.55rem; }
280
+ .vgm-feature-list li {
281
+ display: flex; gap: 0.6rem; align-items: flex-start;
282
+ font-size: 0.88rem; line-height: 1.45;
283
+ }
284
+ .vgm-feature-list li::before {
285
+ content: '';
286
+ flex: 0 0 0.45rem;
287
+ width: 0.45rem; height: 0.45rem;
288
+ border-radius: 50%;
289
+ background: var(--color-primary);
290
+ margin-top: 0.5rem;
291
+ }
292
+
293
+ /* ADVANCED DETAILS */
294
+ details.vgm-advanced {
295
  border: 0;
296
+ padding: 0;
297
  }
298
+ details.vgm-advanced > summary {
299
+ list-style: none;
300
+ cursor: pointer;
301
+ display: inline-flex;
302
+ align-items: center;
303
+ gap: 0.4rem;
304
+ font-size: 0.84rem;
305
+ font-weight: 600;
306
+ color: var(--color-primary);
307
+ padding: 0.4rem 0;
308
  }
309
+ details.vgm-advanced > summary::-webkit-details-marker { display: none; }
310
+ details.vgm-advanced > summary::before {
311
+ content: '+';
312
+ font-family: var(--font-mono);
313
+ font-size: 1rem;
314
+ width: 1rem;
315
+ text-align: center;
316
+ }
317
+ details.vgm-advanced[open] > summary::before { content: '−'; }
318
+ details.vgm-advanced > div { padding-top: var(--space-sm); }
319
+
320
+ /* TUNING SLIDERS (carried over) */
321
  .vgm-tune {
322
+ margin: 0.6rem 0 1.1rem;
323
+ padding-bottom: 0.7rem;
324
+ border-bottom: 1px solid var(--color-border-soft);
325
  }
326
  .vgm-tune:last-of-type { border-bottom: 0; }
327
  .vgm-tune__head {
328
+ display: flex; align-items: baseline; justify-content: space-between;
329
+ gap: 0.6rem; margin-bottom: 0.3rem;
 
 
 
 
 
 
 
330
  }
331
+ .vgm-tune__head label { font-weight: 600; font-size: 0.92rem; }
332
  .vgm-tune__value {
333
+ font-family: var(--font-mono);
334
+ font-size: 0.85rem;
335
+ color: var(--color-text);
336
+ background: var(--color-surface-alt);
337
  padding: 0.1rem 0.5rem;
338
  border-radius: 4px;
339
  min-width: 4ch;
340
  text-align: center;
341
  }
342
+ .vgm-tune input[type="range"] { width: 100%; accent-color: var(--color-primary); margin: 0.15rem 0; }
 
 
 
 
343
  .vgm-tune input[type="range"]:disabled { opacity: 0.45; }
344
  .vgm-tune__anchors {
345
+ display: flex; justify-content: space-between;
346
+ font-size: 0.72rem; color: var(--color-text-subtle);
347
+ letter-spacing: 0.02em; text-transform: uppercase;
 
 
 
348
  margin-top: -0.1rem;
349
  }
350
  .vgm-tune__help {
351
+ display: block; margin-top: 0.4rem;
352
+ font-size: 0.82rem; color: var(--color-text-muted); line-height: 1.4;
 
 
 
353
  }
354
  .vgm-tune__toggle {
355
+ display: inline-flex; align-items: center; gap: 0.4rem;
356
+ font-size: 0.85rem; margin-bottom: 0.3rem;
357
+ }
358
+
359
+ /* PROGRESS PANEL */
360
+ .vgm-progress {
361
+ display: none;
362
+ background: var(--color-surface);
363
+ border: 1px solid var(--color-border);
364
+ border-radius: var(--radius-md);
365
+ padding: var(--space-lg);
366
+ box-shadow: var(--shadow-sm);
367
+ margin-bottom: var(--space-lg);
368
+ }
369
+ .vgm-progress[data-state="running"], .vgm-progress[data-state="error"] { display: block; }
370
+ .vgm-progress h2 { font-family: var(--font-heading); font-size: 1.35rem; margin: 0 0 0.6rem; }
371
+ .vgm-progress progress {
372
+ width: 100%; height: 0.55rem;
373
+ accent-color: var(--color-primary);
374
+ border-radius: 4px;
375
+ }
376
+ .vgm-progress__msg { display: flex; gap: 0.5rem; align-items: baseline; margin-top: 0.6rem; font-size: 0.92rem; }
377
+ .vgm-progress__stage {
378
+ display: inline-block;
379
+ font-size: 0.7rem; font-weight: 700;
380
+ text-transform: uppercase; letter-spacing: 0.07em;
381
+ color: #fff; background: var(--color-primary);
382
+ padding: 0.18rem 0.55rem; border-radius: 3px;
383
+ }
384
+ .vgm-progress__text { color: var(--color-text); }
385
+ .vgm-progress__details {
386
+ display: none;
387
+ margin-top: 0.7rem;
388
+ padding: 0.6rem 0.8rem;
389
+ background: #fff; border: 1px solid #f5b5b5;
390
+ border-radius: 4px;
391
+ font-family: var(--font-mono);
392
+ font-size: 0.8rem; color: #5a0808;
393
+ white-space: pre-wrap; word-break: break-word;
394
+ max-height: 14rem; overflow: auto;
395
+ }
396
+ .vgm-progress[data-state="error"] .vgm-progress__details:not(:empty) { display: block; }
397
+ .vgm-progress__retry {
398
+ display: none;
399
+ margin-top: 0.7rem;
400
+ padding: 0.45rem 0.95rem;
401
+ background: var(--color-primary);
402
+ color: #fff; border: 0; border-radius: 4px;
403
+ font-family: inherit; font-size: 0.9rem; cursor: pointer;
404
+ }
405
+ .vgm-progress[data-state="error"] {
406
+ background: var(--color-error-bg);
407
+ border-color: #f5b5b5;
408
+ }
409
+ .vgm-progress[data-state="error"] .vgm-progress__retry { display: inline-block; }
410
+ .vgm-progress__stages {
411
+ display: flex; gap: 0.35rem; flex-wrap: wrap;
412
+ margin-top: 0.8rem; font-size: 0.72rem;
413
+ }
414
+ .vgm-progress__stages span {
415
+ padding: 0.2rem 0.55rem;
416
+ border: 1px solid var(--color-border);
417
+ border-radius: 999px;
418
+ color: var(--color-text-subtle);
419
+ background: var(--color-surface);
420
+ }
421
+ .vgm-progress__stages span.is-done {
422
+ background: var(--color-surface-alt);
423
+ color: var(--color-text-muted);
424
+ border-color: var(--color-border-soft);
425
+ }
426
+ .vgm-progress__stages span.is-active {
427
+ background: var(--color-primary);
428
+ color: #fff;
429
+ border-color: var(--color-primary);
430
+ }
431
+
432
+ /* REVIEW IFRAME OVERLAY */
433
+ .vgm-review {
434
+ display: none;
435
+ position: fixed; inset: 0;
436
+ background: #fff; z-index: 100;
437
+ flex-direction: column;
438
+ }
439
+ .vgm-review[data-state="open"] { display: flex; }
440
+ .vgm-review__bar {
441
+ flex: 0 0 auto;
442
+ display: flex; align-items: center; gap: 0.75rem;
443
+ padding: 0.5rem 1rem;
444
+ background: #1a1a1a; color: #fff;
445
+ border-bottom: 1px solid #333;
446
+ font-size: 0.9rem;
447
+ }
448
+ .vgm-review__bar strong { font-weight: 600; }
449
+ .vgm-review__bar button {
450
+ background: transparent; border: 1px solid #555; color: #fff;
451
+ padding: 0.3rem 0.7rem; border-radius: 4px;
452
+ font-family: inherit; font-size: 0.85rem; cursor: pointer;
453
  }
454
+ .vgm-review__bar button:hover { background: #2a2a2a; }
455
+ .vgm-review__bar .spacer { flex: 1; }
456
+ .vgm-review__frame { flex: 1 1 auto; width: 100%; border: 0; }
457
  </style>
458
  </head>
459
  <body>
460
  <a class="skip-link" href="#main">Skip to main content</a>
 
 
 
 
461
 
462
+ <div class="vgm-page" id="main" tabindex="-1">
463
+
464
+ <header>
465
+ <span class="vgm-eyebrow">Video Guide Maker</span>
466
+ <h1 class="vgm-title">Turn lectures into <em>study-ready</em> guides.</h1>
467
+ <p class="vgm-tagline">
468
+ Upload a video and its transcript. Get a WCAG-ready HTML guide with the cleanest
469
+ slide frames, aligned narration, and — optionally — Claude-generated section
470
+ titles, key terms, alt-text, and equations you can review and refine before
471
+ publishing.
472
+ </p>
473
+ </header>
474
 
475
  <div class="vgm-progress" data-state="idle" role="status" aria-live="polite" aria-atomic="true">
476
+ <h2>Generating your study guide…</h2>
477
  <progress max="100" value="0"></progress>
478
  <p class="vgm-progress__msg">
479
  <span class="vgm-progress__stage">starting</span>
480
  <span class="vgm-progress__text">Uploading…</span>
481
  </p>
482
+ <div class="vgm-progress__stages" aria-hidden="true">
483
+ <span data-stage="uploading">upload</span>
484
+ <span data-stage="scene_detect">scenes</span>
485
+ <span data-stage="filter">filter</span>
486
+ <span data-stage="transcript">transcript</span>
487
+ <span data-stage="ocr">ocr</span>
488
+ <span data-stage="llm_extract">llm</span>
489
+ <span data-stage="render">render</span>
490
+ </div>
491
  <pre class="vgm-progress__details" aria-label="Error details"></pre>
492
  <button type="button" class="vgm-progress__retry">Try again</button>
493
  </div>
494
 
495
+ <div class="vgm-hero">
 
496
 
497
+ <form id="vgm-form" action="/process" method="post" enctype="multipart/form-data" class="vgm-card" aria-labelledby="form-heading">
498
+ <h2 id="form-heading" class="sr-only" style="position:absolute;left:-9999px;">Generate study guide</h2>
 
 
499
 
500
+ <div class="vgm-section">
501
+ <h3 class="vgm-section__title">1. Upload</h3>
 
 
502
 
503
+ <div class="vgm-field">
504
+ <label for="title">Study guide title</label>
505
+ <input type="text" id="title" name="title" value="Lecture Study Guide" required>
506
+ </div>
507
 
508
+ <div class="vgm-field">
509
+ <label for="video">Video file</label>
510
+ <label class="vgm-drop" data-loaded="false">
511
+ <input type="file" id="video" name="video" accept="video/*" required>
512
+ <div class="vgm-drop__hint">
513
+ <strong>Drop a video here</strong> or click to browse<br>
514
+ <small>MP4, MOV, MKV, WebM, M4V</small>
515
+ </div>
516
+ <div class="vgm-drop__meta" data-meta-for="video"></div>
517
+ </label>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
518
  </div>
 
 
 
 
 
 
 
 
519
 
520
+ <div class="vgm-field">
521
+ <label for="transcript">Transcript</label>
522
+ <label class="vgm-drop" data-loaded="false">
523
+ <input type="file" id="transcript" name="transcript" accept=".vtt,.srt,text/vtt,application/x-subrip" required>
524
+ <div class="vgm-drop__hint">
525
+ <strong>Drop a VTT or SRT file</strong> or click to browse<br>
526
+ <small>Cue timestamps drive segment alignment</small>
527
+ </div>
528
+ <div class="vgm-drop__meta" data-meta-for="transcript"></div>
529
+ </label>
530
  </div>
 
 
 
 
 
 
531
  </div>
532
 
533
+ <div class="vgm-section">
534
+ <h3 class="vgm-section__title">2. AI assist (optional)</h3>
535
+ <p class="vgm-section__intro">
536
+ Paste an Anthropic API key to let Claude generate the fields below. Without a key the
537
+ pipeline still runs — you'll just get the OCR-derived scaffold with placeholders.
538
+ Toggle off any feature you don't want to pay for.
539
+ </p>
540
+
541
+ <div class="vgm-field">
542
+ <label for="anthropic_api_key">Anthropic API key</label>
543
+ <input type="password" id="anthropic_api_key" name="anthropic_api_key"
544
+ autocomplete="off" spellcheck="false"
545
+ placeholder="sk-ant-…">
546
+ <small>Used once for this run. Sent over HTTPS, held in memory only, never written to disk or logs.</small>
547
+ </div>
548
+
549
+ <div class="vgm-llm-grid">
550
+ <label class="vgm-llm-toggle">
551
+ <input type="checkbox" name="llm_section_title" value="true" checked>
552
+ <span class="vgm-llm-toggle__body">
553
+ <strong>Section titles</strong>
554
+ <small>"Combining AI models" instead of "Segment 3 — 4:25"</small>
555
+ </span>
556
+ </label>
557
+ <label class="vgm-llm-toggle">
558
+ <input type="checkbox" name="llm_alt_text" value="true" checked>
559
+ <span class="vgm-llm-toggle__body">
560
+ <strong>Alt-text drafts</strong>
561
+ <small>WCAG-purpose descriptions, not OCR transcripts</small>
562
+ </span>
563
+ </label>
564
+ <label class="vgm-llm-toggle">
565
+ <input type="checkbox" name="llm_key_terms" value="true" checked>
566
+ <span class="vgm-llm-toggle__body">
567
+ <strong>Key terms</strong>
568
+ <small>Term + concise definition per segment</small>
569
+ </span>
570
+ </label>
571
+ <label class="vgm-llm-toggle">
572
+ <input type="checkbox" name="llm_math_equations" value="true" checked>
573
+ <span class="vgm-llm-toggle__body">
574
+ <strong>Math equations</strong>
575
+ <small>LaTeX from on-screen formulas, with screen-reader labels</small>
576
+ </span>
577
+ </label>
578
  </div>
 
 
 
 
 
 
 
 
 
579
  </div>
580
 
581
+ <div class="vgm-section">
582
+ <h3 class="vgm-section__title">3. Output</h3>
583
+ <div class="vgm-formats">
584
+ <label class="vgm-fmt">
585
+ <span class="vgm-fmt__head"><input type="radio" name="format" value="review" checked>Review &amp; edit</span>
586
+ <small>Open the in-browser editor to refine before publishing.</small>
587
+ </label>
588
+ <label class="vgm-fmt">
589
+ <span class="vgm-fmt__head"><input type="radio" name="format" value="single">Single HTML</span>
590
+ <small>Self-contained file with all images inlined.</small>
591
+ </label>
592
+ <label class="vgm-fmt">
593
+ <span class="vgm-fmt__head"><input type="radio" name="format" value="zip">Zip bundle</span>
594
+ <small>HTML plus a static folder, ready to host.</small>
595
+ </label>
596
  </div>
 
 
 
 
 
 
 
597
  </div>
598
 
599
+ <div class="vgm-section">
600
+ <details class="vgm-advanced">
601
+ <summary>Advanced settings</summary>
602
+ <div>
603
+
604
+ <div class="vgm-tune">
605
+ <div class="vgm-tune__head">
606
+ <label for="threshold">Scene-change sensitivity</label>
607
+ <output for="threshold" class="vgm-tune__value">27</output>
608
+ </div>
609
+ <input type="range" id="threshold" name="threshold" value="27" min="5" max="60" step="1">
610
+ <div class="vgm-tune__anchors"><span>More scenes</span><span>Fewer scenes</span></div>
611
+ <small class="vgm-tune__help">
612
+ How visually different two frames must be before they count as a new scene.
613
+ Lower (5–15) catches every animation step; higher (35+) only major slide changes.
614
+ </small>
615
+ </div>
616
+
617
+ <div class="vgm-tune">
618
+ <div class="vgm-tune__head">
619
+ <label for="min_gap">Minimum gap between scenes</label>
620
+ <output for="min_gap" class="vgm-tune__value">off</output>
621
+ </div>
622
+ <input type="range" id="min_gap" name="min_gap" value="0" min="0" max="60" step="0.5">
623
+ <div class="vgm-tune__anchors"><span>Off</span><span>60 s</span></div>
624
+ <small class="vgm-tune__help">
625
+ Drop new scenes that arrive less than this many seconds after the previous kept one.
626
+ Useful for slides that build up gradually — try ~5 s to keep just the final state.
627
+ </small>
628
+ </div>
629
+
630
+ <div class="vgm-tune">
631
+ <div class="vgm-tune__head">
632
+ <label for="max_frames">Max frames</label>
633
+ <output for="max_frames" class="vgm-tune__value">unlimited</output>
634
+ </div>
635
+ <label class="vgm-tune__toggle">
636
+ <input type="checkbox" id="max_frames_enabled"> Cap total frames
637
+ </label>
638
+ <input type="range" id="max_frames" min="1" max="200" step="1" value="50" disabled>
639
+ <div class="vgm-tune__anchors"><span>1</span><span>200</span></div>
640
+ <small class="vgm-tune__help">
641
+ Hard limit on the number of frames in the guide. Useful for very long lectures
642
+ or to bound LLM extraction cost. Frames are evenly distributed across the video.
643
+ </small>
644
+ </div>
645
+
646
+ <div class="vgm-tune">
647
+ <div class="vgm-tune__head">
648
+ <label for="face_threshold">Instructor-frame face threshold</label>
649
+ <output for="face_threshold" class="vgm-tune__value">0.12</output>
650
+ </div>
651
+ <input type="range" id="face_threshold" name="face_threshold" value="0.12" min="0.02" max="0.6" step="0.01">
652
+ <div class="vgm-tune__anchors"><span>Strict (drop more)</span><span>Lenient (keep more)</span></div>
653
+ <small class="vgm-tune__help">
654
+ Fraction of the frame a detected face must occupy before we drop it as
655
+ instructor-only. Lower = more aggressive; raise it if your slides have an
656
+ inset webcam you want to keep.
657
+ </small>
658
+ </div>
659
+
660
+ <div class="vgm-tune">
661
+ <label class="vgm-tune__toggle">
662
+ <input type="checkbox" name="skip_ocr" value="true"> Skip OCR
663
+ </label>
664
+ <small class="vgm-tune__help">
665
+ Faster; on-screen text and the heuristic LaTeX wrapping in the guide will be empty.
666
+ Doesn't affect LLM extraction (which reads the frame image directly).
667
+ </small>
668
+ </div>
669
+
670
+ <div class="vgm-tune">
671
+ <div class="vgm-tune__head">
672
+ <label for="lang">Document language</label>
673
+ <output for="lang" class="vgm-tune__value">en</output>
674
+ </div>
675
+ <input type="text" id="lang" name="lang" value="en" pattern="[a-zA-Z\-]+" maxlength="12" style="width: 6rem;">
676
+ <small class="vgm-tune__help">
677
+ BCP-47 code (<code>en</code>, <code>en-US</code>, <code>fr</code>…).
678
+ Sets the <code>lang</code> attribute on the generated HTML and selects
679
+ Tesseract language packs when available.
680
+ </small>
681
+ </div>
682
+
683
+ </div>
684
+ </details>
685
  </div>
686
 
687
+ <div class="vgm-submit-row">
688
+ <span class="vgm-submit-row__hint">Processing usually takes ~30 s per minute of video.</span>
689
+ <button type="submit" class="vgm-cta">Generate study guide →</button>
 
 
 
 
 
 
 
 
 
690
  </div>
691
+
692
+ </form>
693
+
694
+ <aside class="vgm-side">
695
+ <div class="vgm-card">
696
+ <h3>What you'll get</h3>
697
+ <ul class="vgm-feature-list">
698
+ <li>One scene-aligned segment per slide with the cleanest available frame.</li>
699
+ <li>Three candidate frames per segment — pick the best in the editor.</li>
700
+ <li>Narration paragraphed and aligned to each visual.</li>
701
+ <li>WCAG 2.2 scaffold: proper heading hierarchy, lang, skip links, alt-text drafts.</li>
702
+ <li>Inline editor for human review — fix titles, alt text, terms, equations.</li>
703
+ </ul>
704
+ </div>
705
+ <div class="vgm-card">
706
+ <h3>Tips</h3>
707
+ <ul class="vgm-feature-list">
708
+ <li><strong>Save money:</strong> uncheck AI features above to skip Claude entirely.</li>
709
+ <li><strong>Progressive-reveal slides:</strong> set Min-gap ≈ 5 s in advanced settings.</li>
710
+ <li><strong>Inset webcam stays:</strong> raise the face-threshold; instructor-only frames still drop.</li>
711
+ <li><strong>Long lectures:</strong> cap Max frames to bound runtime and API cost.</li>
712
+ </ul>
713
+ </div>
714
+ </aside>
715
+
716
+ </div>
717
+ </div>
718
 
719
  <div class="vgm-review" data-state="closed" role="dialog" aria-label="Study guide editor">
720
  <div class="vgm-review__bar">
 
736
  const textEl = panel.querySelector('.vgm-progress__text');
737
  const detailsEl = panel.querySelector('.vgm-progress__details');
738
  const retryBtn = panel.querySelector('.vgm-progress__retry');
739
+ const stagePillsEl = panel.querySelector('.vgm-progress__stages');
740
  const apiKeyEl = form.querySelector('#anthropic_api_key');
741
  const titleEl = form.querySelector('#title');
742
  const reviewEl = document.querySelector('.vgm-review');
 
746
  const reviewOpenBtn = reviewEl.querySelector('.vgm-review__open');
747
  let lastResultUrl = null;
748
 
749
+ const STAGE_ORDER = ['uploading', 'scene_detect', 'filter', 'transcript', 'ocr', 'llm_extract', 'render'];
750
+ function setStagePill(stage) {
751
+ const idx = STAGE_ORDER.indexOf(stage);
752
+ stagePillsEl.querySelectorAll('span').forEach(el => {
753
+ const i = STAGE_ORDER.indexOf(el.getAttribute('data-stage'));
754
+ el.classList.toggle('is-active', i === idx);
755
+ el.classList.toggle('is-done', i >= 0 && idx > i);
756
+ });
757
+ }
758
+
759
  function openReview(resultUrl, title) {
760
  lastResultUrl = resultUrl;
761
  reviewTitleEl.textContent = title ? '— ' + title : '';
 
773
  if (lastResultUrl) window.open(lastResultUrl, '_blank', 'noopener');
774
  });
775
 
776
+ // Drop zones — show file metadata + drag styling.
777
+ function fmtBytes(n) {
778
+ if (n < 1024) return n + ' B';
779
+ if (n < 1024 * 1024) return (n / 1024).toFixed(1) + ' KB';
780
+ if (n < 1024 * 1024 * 1024) return (n / (1024 * 1024)).toFixed(1) + ' MB';
781
+ return (n / (1024 * 1024 * 1024)).toFixed(2) + ' GB';
782
+ }
783
+ function fmtDuration(s) {
784
+ if (!s || !isFinite(s)) return '—';
785
+ const h = Math.floor(s / 3600);
786
+ const m = Math.floor((s % 3600) / 60);
787
+ const sec = Math.round(s % 60);
788
+ return h > 0 ? `${h}:${String(m).padStart(2,'0')}:${String(sec).padStart(2,'0')}`
789
+ : `${m}:${String(sec).padStart(2,'0')}`;
790
+ }
791
+ async function readVideoMeta(file) {
792
+ return new Promise(resolve => {
793
+ const v = document.createElement('video');
794
+ v.preload = 'metadata';
795
+ v.onloadedmetadata = () => {
796
+ const d = v.duration;
797
+ URL.revokeObjectURL(v.src);
798
+ resolve({ duration: d, width: v.videoWidth, height: v.videoHeight });
799
+ };
800
+ v.onerror = () => { URL.revokeObjectURL(v.src); resolve({}); };
801
+ v.src = URL.createObjectURL(file);
802
+ });
803
+ }
804
+ async function readTranscriptPreview(file) {
805
+ try {
806
+ const text = await file.slice(0, 8 * 1024).text();
807
+ const cues = (text.match(/-->/g) || []).length;
808
+ const firstCue = (text.match(/(\d{1,2}:\d{2}:\d{2}[.,]\d{3})\s*-->\s*(\d{1,2}:\d{2}:\d{2}[.,]\d{3})/) || [])[0];
809
+ return { cues, firstCue: firstCue || '' };
810
+ } catch (_) { return {}; }
811
+ }
812
+ function setMeta(dropEl, label, rows) {
813
+ const meta = dropEl.querySelector('.vgm-drop__meta');
814
+ const rowsHTML = rows.filter(([k, v]) => v).map(([k, v]) =>
815
+ `<dt>${k}</dt><dd>${v}</dd>`
816
+ ).join('');
817
+ meta.innerHTML = `<strong>${label}</strong>${rowsHTML ? '<dl>' + rowsHTML + '</dl>' : ''}`;
818
+ dropEl.setAttribute('data-loaded', 'true');
819
+ }
820
+ function bindDropZone(input, kind) {
821
+ const drop = input.closest('.vgm-drop');
822
+ ['dragenter', 'dragover'].forEach(evt =>
823
+ drop.addEventListener(evt, e => { e.preventDefault(); drop.classList.add('is-dragover'); })
824
+ );
825
+ ['dragleave', 'drop'].forEach(evt =>
826
+ drop.addEventListener(evt, e => { drop.classList.remove('is-dragover'); })
827
+ );
828
+ input.addEventListener('change', async () => {
829
+ const f = input.files && input.files[0];
830
+ if (!f) { drop.setAttribute('data-loaded', 'false'); return; }
831
+ if (kind === 'video') {
832
+ setMeta(drop, f.name, [['Size', fmtBytes(f.size)], ['Type', f.type || '—']]);
833
+ const meta = await readVideoMeta(f);
834
+ if (meta.duration) {
835
+ const dim = (meta.width && meta.height) ? `${meta.width}×${meta.height}` : '';
836
+ setMeta(drop, f.name, [
837
+ ['Size', fmtBytes(f.size)],
838
+ ['Duration', fmtDuration(meta.duration)],
839
+ ['Resolution', dim],
840
+ ]);
841
+ }
842
+ } else {
843
+ setMeta(drop, f.name, [['Size', fmtBytes(f.size)]]);
844
+ const t = await readTranscriptPreview(f);
845
+ if (t.cues != null) {
846
+ setMeta(drop, f.name, [
847
+ ['Size', fmtBytes(f.size)],
848
+ ['Cues (approx.)', String(t.cues)],
849
+ ['First cue', t.firstCue || '—'],
850
+ ]);
851
+ }
852
+ }
853
+ });
854
+ }
855
+ bindDropZone(form.querySelector('#video'), 'video');
856
+ bindDropZone(form.querySelector('#transcript'), 'transcript');
857
+
858
+ // Sliders: keep <output> in sync.
859
  function fmtMinGap(v) { return parseFloat(v) <= 0 ? 'off' : (parseFloat(v) + ' s'); }
860
  function fmtIdentity(v) { return v; }
861
  const formatters = { min_gap: fmtMinGap };
 
867
  slider.addEventListener('input', sync);
868
  sync();
869
  });
 
870
  form.querySelectorAll('input[type="text"][id]').forEach(input => {
871
  const out = form.querySelector('output[for="' + input.id + '"]');
872
  if (!out) return;
 
875
  sync();
876
  });
877
 
878
+ // Max-frames cap toggle.
 
879
  const maxFramesToggle = form.querySelector('#max_frames_enabled');
880
  const maxFramesSlider = form.querySelector('#max_frames');
881
  const maxFramesOut = form.querySelector('output[for="max_frames"]');
 
899
  function setProgress(state, percent, stage, message) {
900
  panel.setAttribute('data-state', state);
901
  if (typeof percent === 'number') bar.value = percent;
902
+ if (stage) {
903
+ stageEl.textContent = stage.replace('_', ' ');
904
+ setStagePill(stage);
905
+ }
906
  if (message) textEl.textContent = message;
907
  }
 
908
  function showError(short, detail) {
909
  setProgress('error', 100, 'error', short);
910
  detailsEl.textContent = detail || '';
911
  }
 
912
  function resetForRetry() {
913
  detailsEl.textContent = '';
914
  panel.setAttribute('data-state', 'idle');
915
  form.style.display = '';
916
  }
 
917
  retryBtn.addEventListener('click', resetForRetry);
918
 
919
  function postJob(fd) {
 
942
  const v = fd.get(k);
943
  if (typeof v === 'string' && v.trim() === '') fd.delete(k);
944
  });
945
+ // Unchecked LLM toggles must explicitly send "false" so the server doesn't apply its default-on.
946
+ ['llm_alt_text', 'llm_key_terms', 'llm_math_equations', 'llm_section_title'].forEach(k => {
947
+ if (!fd.has(k)) fd.set(k, 'false');
948
+ });
949
  if (apiKeyEl) apiKeyEl.value = '';
950
  const submittedTitle = (titleEl && titleEl.value) || 'Study Guide';
951
  const formatRadio = form.querySelector('input[name="format"]:checked');
 
967
  try {
968
  const parsed = JSON.parse(detail);
969
  detail = parsed.detail || parsed.message || detail;
970
+ } catch (_) {}
971
  showError(`Server rejected the upload (HTTP ${xhr.status})`, detail);
972
  return;
973
  }
 
991
  setProgress('running', data.percent, data.stage, data.message);
992
  } else if (data.type === 'done') {
993
  es.close();
994
+ setProgress('running', 100, 'render', 'Loading editor…');
995
  if (format === 'review') {
996
  openReview(result_url, submittedTitle);
 
997
  panel.setAttribute('data-state', 'idle');
998
  form.style.display = '';
999
  } else {
 
1002
  document.body.appendChild(a);
1003
  a.click();
1004
  a.remove();
1005
+ setProgress('running', 100, 'render', 'Downloaded. Generate another?');
1006
  setTimeout(resetForRetry, 2500);
1007
  }
1008
  } else if (data.type === 'error') {
 
1011
  }
1012
  };
1013
  es.onerror = () => {
 
 
1014
  if (es.readyState === EventSource.CLOSED) {
1015
  showError('Lost connection to server.',
1016
  `Last progress message: ${lastStageMsg}\n` +