Streamline Llama processor

src/aiprocess.py

@@ -17,11 +17,12 @@ SYSTEM_PROMPT = (
     "Rules:\n"
     "- Remove filler words (um/uh/like), false starts, and self-corrections.\n"
     "- Keep meaning, facts, and intent.\n"
+    "- Preserve greetings and salutations.\n"
     "- Prefer concise sentences.\n"
     "- Do not add new info.\n"
     "- Output ONLY the cleaned text, no commentary.\n\n"
     "Examples:\n"
-    "   - \"schedule that for 5 PM, I mean 4 PM\" -> \"schedule that for 4 PM\"\n"
+    "   - \"Hey, schedule that for 5 PM, I mean 4 PM\" -> \"Hey, schedule that for 4 PM\"\n"
     "   - \"let's ask Bob, I mean Janice, let's ask Janice\" -> \"let's ask Janice\"\n"
 )
 
@@ -32,36 +33,27 @@ MODEL_URL = (
 )
 MODEL_DIR = Path.home() / ".cache" / "lel" / "models"
 MODEL_PATH = MODEL_DIR / MODEL_NAME
-LLM_LANGUAGE = "en"
-
-
-@dataclass
-class LLMConfig:
-    model_path: Path
-    n_ctx: int = 4096
-    verbose: bool = False
 
 
 class LlamaProcessor:
-    def __init__(self, cfg: LLMConfig):
-        self.cfg = cfg
-        if not cfg.verbose:
+    def __init__(self, verbose=False):
+        ensure_model()
+        if not verbose:
             os.environ.setdefault("LLAMA_CPP_LOG_LEVEL", "ERROR")
             os.environ.setdefault("LLAMA_LOG_LEVEL", "ERROR")
-        self._log_callback = _llama_log_callback_factory(cfg.verbose)
+        self._log_callback = _llama_log_callback_factory(verbose)
         llama_cpp_lib.llama_log_set(cast(Any, self._log_callback), ctypes.c_void_p())
         os.environ.setdefault("LLAMA_CPP_LOG_PREFIX", "llama")
         os.environ.setdefault("LLAMA_CPP_LOG_PREFIX_SEPARATOR", "::")
         self.client = Llama(
-            model_path=str(cfg.model_path),
-            n_ctx=cfg.n_ctx,
-            verbose=cfg.verbose,
+            model_path=str(MODEL_PATH),
+            n_ctx=4096,
+            verbose=verbose,
         )
 
-    def process(self, text: str) -> str:
+    def process(self, text: str, lang: str = "en") -> str:
         user_content = f"<transcript>{text}</transcript>"
-        if LLM_LANGUAGE:
-            user_content = f"<language>{LLM_LANGUAGE}</language>\n{user_content}"
+        user_content = f"<language>{lang}</language>\n{user_content}"
         response = self.client.create_chat_completion(
             messages=[
                 {"role": "system", "content": SYSTEM_PROMPT},
@@ -72,12 +64,7 @@ class LlamaProcessor:
         return _extract_chat_text(response)
 
 
-def build_processor(verbose: bool = False) -> LlamaProcessor:
-    model_path = ensure_model()
-    return LlamaProcessor(LLMConfig(model_path=model_path, verbose=verbose))
-
-
-def ensure_model() -> Path:
+def ensure_model():
     if MODEL_PATH.exists():
         return MODEL_PATH
     MODEL_DIR.mkdir(parents=True, exist_ok=True)
@@ -111,7 +98,6 @@ def ensure_model() -> Path:
         except Exception:
             pass
         raise
-    return MODEL_PATH
 
 
 def _extract_chat_text(payload: Any) -> str:

src/leld.py

@@ -14,7 +14,7 @@ from faster_whisper import WhisperModel
 
 from config import Config, load, redacted_dict
 from recorder import start_recording, stop_recording
-from aiprocess import build_processor
+from aiprocess import LlamaProcessor
 from desktop import get_desktop_adapter
 
 
@@ -38,7 +38,7 @@ def _compute_type(device: str) -> str:
 
 
 class Daemon:
-    def __init__(self, cfg: Config, desktop, *, llama_verbose: bool = False):
+    def __init__(self, cfg: Config, desktop, *, verbose: bool = False):
         self.cfg = cfg
         self.desktop = desktop
         self.lock = threading.Lock()
@@ -51,7 +51,7 @@ class Daemon:
             device=cfg.stt.get("device", "cpu"),
             compute_type=_compute_type(cfg.stt.get("device", "cpu")),
         )
-        self.ai_processor = build_processor(verbose=llama_verbose)
+        self.ai_processor = LlamaProcessor(verbose=verbose)
 
     def set_state(self, state: str):
         with self.lock:
@@ -210,7 +210,6 @@ def main():
     cfg = load(args.config)
     _lock_single_instance()
 
-    logging.info("ready")
     logging.info("hotkey: %s", cfg.daemon.get("hotkey", ""))
     logging.info("config (%s):\n%s", args.config or str(Path.home() / ".config" / "lel" / "config.json"), json.dumps(redacted_dict(cfg), indent=2))
 
@@ -218,7 +217,7 @@ def main():
         logging.getLogger().setLevel(logging.DEBUG)
     desktop = get_desktop_adapter()
     try:
-        daemon = Daemon(cfg, desktop, llama_verbose=args.verbose)
+        daemon = Daemon(cfg, desktop, verbose=args.verbose)
     except Exception as exc:
         logging.error("startup failed: %s", exc)
         raise SystemExit(1)
@@ -238,6 +237,7 @@ def main():
         cfg.daemon.get("hotkey", ""),
         lambda: logging.info("hotkey pressed (dry-run)") if args.dry_run else daemon.toggle(),
     )
+    logging.info("ready")
     desktop.run_tray(daemon.get_state, daemon._quit)