update new commands

.gitignore

@@ -9,4 +9,6 @@ output/*
 # Python cache
 __pycache__
 *.pyc
-.pytest_cache/
+.pytest_cache/
+
+local-run.sh

compile_meeting.py

@@ -51,6 +51,12 @@ Rules:
 - Reorganize by TOPIC, not by conversation order. Merge new info into the right
   existing section rather than appending chronologically.
 - Anchor concrete items to the [mm:ss] where they were said/shown.
+- ASR vigilance on TERMS, especially acronyms and product/tool names: this
+  transcript is machine-transcribed, so a term that reads oddly or makes no sense
+  in context is likely a mis-hearing (a slightly-off acronym, a homophone, a
+  split or merged word). Flag it like "(heard: X — likely Y?)", using context to
+  infer the intended term; never silently propagate a nonsensical token, and
+  never silently "correct" a term you are unsure about.
 - If something is unclear or only partially stated, mark it (e.g. "(unclear)")
   rather than guessing."""
 
@@ -101,6 +107,19 @@ def default_output(transcript, kind):
     return transcript.parent / f"{stem}_{kind}.md"
 
 
+def content_tokens(content):
+    """Estimate input tokens of a chat content (str or multimodal list)."""
+    if isinstance(content, str):
+        return estimate_tokens(content)
+    total = 0
+    for part in content:
+        if part.get("type") == "text":
+            total += estimate_tokens(part.get("text", ""))
+        else:  # image_url etc. — rough per-image vision-token allowance
+            total += 800
+    return total
+
+
 def parse_windows(path, window_tokens):
     """Split the enhanced transcript into windows of ~window_tokens, packing
     blank-line-separated blocks whole. Each window keeps the frame refs that
@@ -208,7 +227,8 @@ def main():
     p.add_argument("--api-key", default="local")
     p.add_argument("--frames", choices=["ondemand", "window", "none"], default="ondemand")
     p.add_argument("--window-tokens", type=int, default=3500, help="transcript tokens per refine step (default 3500)")
-    p.add_argument("--max-tokens", type=int, default=8192, help="generation cap; must fit the growing doc (default 8192)")
+    p.add_argument("--max-tokens", type=int, default=8192, help="upper bound on generation; auto-capped to fit --ctx (default 8192)")
+    p.add_argument("--ctx", type=int, default=16384, help="model context window; output is auto-capped so input+output fit (default 16384)")
     p.add_argument("--max-image-side", type=int, default=1280, help="downscale frames to this max side (0=off)")
     p.add_argument("--temperature", type=float, default=0.2)
     p.add_argument("--checkpoint", type=Path, help="write the running doc here after each window (resumable progress)")
@@ -260,9 +280,13 @@ def main():
             content = text
             log(f"  window {wi}/{len(windows)}: text-only")
 
-        doc = call(client, args.model,
-                   REFINE_SYS.format(instruction=args.instruction, rules=GROUNDING),
-                   content, args.temperature, args.max_tokens)
+        sys_txt = REFINE_SYS.format(instruction=args.instruction, rules=GROUNDING)
+        in_tok = content_tokens(content) + estimate_tokens(sys_txt)
+        out_budget = max(512, min(args.max_tokens, args.ctx - in_tok - 256))
+        if in_tok > args.ctx - 512:
+            log(f"  WARNING: doc+window ~{in_tok} tok ≥ ctx {args.ctx}; output will truncate — "
+                f"use a 32k profile (qwen14b-gguf) or lower --window-tokens for the full training")
+        doc = call(client, args.model, sys_txt, content, args.temperature, out_budget)
 
         if args.checkpoint:
             args.checkpoint.write_text(doc + "\n")

summarize_meeting.py

@@ -54,7 +54,11 @@ Rules you must follow:
 - The transcript is machine-generated and may contain ASR errors; prefer the
   most consistent reading across the whole transcript over any single garbled
   line, and flag a name/term you are unsure about rather than normalizing it
-  silently."""
+  silently.
+- Pay special attention to acronyms and product/tool names: a slightly-off
+  acronym or a homophone that makes no sense in context is almost certainly an
+  ASR mis-hearing — flag it as "(heard: X — likely Y?)" using context to infer
+  the intended term, rather than repeating the nonsensical form."""
 
 MAP_SYSTEM = """\
 You are condensing ONE chunk of a longer meeting transcript.