changed describe.py a bit

bot.py

@@ -93,7 +93,7 @@ QUIET_COOLDOWN = 600  # seconds of silence before she's allowed to send a random
 MEDIA_MIMES = ("image/png", "image/jpeg", "image/webp", "image/heic", "image/heif",
                "audio/", "video/", "application/pdf", "text/plain")  # what gemini can actually read
 MEDIA_MAX_BYTES = 10 * 1024 * 1024  # total per message, gemini caps inline request size (~20mb and base64 adds a third)
-MEDIA_SAVE_MAX_BYTES = 8 * 1024 * 1024  # per file we keep, stays under discord's 10mb upload limit so she can send it back
+MEDIA_SAVE_MAX_BYTES = 19 * 1024 * 1024  # per file we keep, stays under discord's 10mb upload limit so she can send it back
 
 def log_error(text: str):
     with ERROR_FILE.open("a", encoding="utf-8") as f:

describe.py

@@ -34,6 +34,10 @@ PROMPT = ("this is a {kind} from a discord server, its name is \"{hint}\" (the n
           "the name above can help), what they're doing, and any text written in the image. "
           "don't guess when to use it, no preamble, no quotes.")
 
+# only for media, the bot sends [media: name] so the file name is what she sees, a hash tells her nothing
+NAME_PROMPT = (" answer in exactly two lines. line 1: a short file name for it, 1 to 3 lowercase words joined by underscores, no extension "
+               "(the character's name if you recognize them, otherwise the main subject). line 2: the description.")
+
 
 def model_from(script: str) -> str:
     # same trick as colab_setup.py, read the model name straight out of main.py / gemini.py so they can't drift apart
@@ -92,7 +96,7 @@ def video_frame(path: Path, dest: Path) -> bool:
 def describe(model: str, png: Path, prompt: str) -> str:
     import ollama
     r = ollama.chat(model=model, messages=[{"role": "user", "content": prompt, "images": [str(png)]}], options={"temperature": 0.2})
-    return clean(r["message"]["content"])
+    return strip_think(r["message"]["content"])
 
 
 def describe_gemini(model: str, png: Path, prompt: str) -> str:
@@ -109,18 +113,52 @@ def describe_gemini(model: str, png: Path, prompt: str) -> str:
         config=types.GenerateContentConfig(safety_settings=safety, temperature=0.2),
     )
     time.sleep(4)  # free tier rate limit, rerun the script if some fail, it only does the empty ones
-    return clean(r.text or "")
+    return strip_think(r.text or "")
+
+
+def strip_think(text: str) -> str:
+    return re.sub(r'<think>.*?</think>', '', text, flags=re.DOTALL)
 
 
 def clean(text: str) -> str:
-    text = re.sub(r'<think>.*?</think>', '', text, flags=re.DOTALL)
-    return " ".join(text.split()).strip(' "')
+    return " ".join(strip_think(text).split()).strip(' "')
 
 
-def save_description(kind: str, name: str, text: str, force: bool):
+def split_name(text: str) -> tuple[str, str]:
+    # media answers come back as "name\ndescription", no name (model ignored the format) = keep the file as it is
+    lines = [l for l in strip_think(text).splitlines() if l.strip()]
+    if len(lines) < 2:
+        return "", clean(text)
+    name = re.sub(r"[^a-z0-9]+", "_", lines[0].lower()).strip("_")[:40]
+    return name, clean(" ".join(lines[1:]))
+
+
+def rename_media(old: str, name: str) -> str:
+    # keeps the extension, adds _2 _3 if the name is taken. media.txt gets the new name too or the bot forgets the file on restart
+    suffix = Path(old).suffix
+    new = f"{name}{suffix}"
+    n = 2
+    while (MEDIA_DIR / new).exists() and new != old:
+        new = f"{name}_{n}{suffix}"
+        n += 1
+    if new == old:
+        return old
+    (MEDIA_DIR / old).rename(MEDIA_DIR / new)
+    log = MEDIA_DIR / "media.txt"
+    if log.exists():
+        lines = log.read_text(encoding="utf-8").splitlines()
+        lines = [l[:-len(old)] + new if l.endswith(f"saved media: {old}") else l for l in lines]
+        log.write_text("\n".join(lines) + "\n", encoding="utf-8")
+    return new
+
+
+def save_description(kind: str, name: str, text: str, force: bool, new_name: str = None):
     # read it fresh every time, the bot adds new entries to this file while we run
     data = json.loads(REACTIONS_FILE.read_text(encoding="utf-8"))
     if name in data.get(kind, {}) and (force or not data[kind][name]):
+        if new_name and new_name != name:
+            data[kind][new_name] = data[kind].pop(name)
+            name = new_name
         data[kind][name] = text
         REACTIONS_FILE.write_text(json.dumps(data, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
 
@@ -161,14 +199,23 @@ def main():
                 try:
                     hint = re.sub(r"^\d+_", "", Path(name).stem) if kind == "media" else name  # media is saved as id_filename
                     prompt = PROMPT.format(kind=kind.rstrip("s") if kind != "media" else "media file", hint=hint)
+                    if kind == "media":
+                        prompt += NAME_PROMPT
                     text = (describe_gemini if args.gemini else describe)(model, png, prompt)
                 except Exception as e:
                     print(f"skip {kind}/{name}: describe failed: {e}")
                     continue
+                new_name = None
+                if kind == "media":
+                    new_name, text = split_name(text)
+                else:
+                    text = clean(text)
                 if not text:
                     continue
-                save_description(kind, name, text, args.all)
-                print(f"{kind}/{name}: {text}")
+                if new_name:
+                    new_name = rename_media(name, new_name)
+                save_description(kind, name, text, args.all, new_name)
+                print(f"{kind}/{name}{f' -> {new_name}' if new_name and new_name != name else ''}: {text}")
 
 
 if __name__ == "__main__":