npcpy 1.2.34__py3-none-any.whl → 1.2.36__py3-none-any.whl

npcpy/gen/ocr.py

@@ -0,0 +1,187 @@
+"""
+Utilities for running DeepSeek OCR (via Unsloth) to turn images into text.
+
+This is intentionally lightweight: the model is only downloaded/loaded when
+`DeepSeekOCR.run` is called. You can point `model_id` at a local path or a
+Hugging Face repo ID; we default to the public `unsloth/DeepSeek-OCR`.
+"""
+
+from __future__ import annotations
+
+import os
+import tempfile
+from dataclasses import dataclass
+from typing import Optional, Union
+
+try:
+    from PIL import Image
+except ImportError:
+    Image = None  # Delayed import for lightweight environments
+
+ImageInput = Union[str, bytes, "Image.Image"]
+
+
+@dataclass
+class DeepSeekOCR:
+    """Lazy loader/wrapper around the Unsloth DeepSeek OCR vision model."""
+
+    model_id: str = "unsloth/DeepSeek-OCR"
+    local_dir: str = os.path.expanduser("~/.npcsh/models/deepseek_ocr")
+    load_in_4bit: bool = False
+    base_size: int = 1024
+    image_size: int = 640
+    crop_mode: bool = True
+
+    def __post_init__(self) -> None:
+        self._model = None
+        self._tokenizer = None
+
+    def _ensure_weights(self) -> str:
+        """Download weights if they are not already on-disk."""
+        if os.path.isdir(self.local_dir) and os.listdir(self.local_dir):
+            return self.local_dir
+
+        os.makedirs(self.local_dir, exist_ok=True)
+        try:
+            from huggingface_hub import snapshot_download
+        except ImportError as exc:
+            raise ImportError(
+                "huggingface_hub is required to download DeepSeek OCR weights. "
+                "Install with `pip install huggingface_hub` or pre-download manually."
+            ) from exc
+
+        snapshot_download(self.model_id, local_dir=self.local_dir)
+        return self.local_dir
+
+    def _load_model(self) -> None:
+        """Load the Unsloth vision model once (lazy)."""
+        if self._model is not None and self._tokenizer is not None:
+            return
+
+        weights_dir = self._ensure_weights()
+        os.environ.setdefault("UNSLOTH_WARN_UNINITIALIZED", "0")
+
+        try:
+            from unsloth import FastVisionModel
+            from transformers import AutoModel
+        except ImportError as exc:
+            raise ImportError(
+                "unsloth and transformers are required to run DeepSeek OCR. "
+                "Install with `pip install unsloth transformers` (and bitsandbytes if using 4bit)."
+            ) from exc
+
+        self._model, self._tokenizer = FastVisionModel.from_pretrained(
+            weights_dir,
+            load_in_4bit=self.load_in_4bit,
+            auto_model=AutoModel,
+            trust_remote_code=True,
+            unsloth_force_compile=True,
+            use_gradient_checkpointing="unsloth",
+        )
+
+    def _prepare_image_file(self, image: ImageInput) -> tuple[str, bool]:
+        """Normalize various image inputs to a file path and say if we should clean it up."""
+        if isinstance(image, str):
+            if not os.path.exists(image):
+                raise FileNotFoundError(f"Image path does not exist: {image}")
+            return image, False
+
+        if Image is None:
+            raise ImportError("Pillow is required for OCR image handling. Install with `pip install pillow`.")
+
+        if isinstance(image, bytes):
+            import io
+
+            pil = Image.open(io.BytesIO(image)).convert("RGB")
+        elif isinstance(image, Image.Image):
+            pil = image.convert("RGB")
+        else:
+            raise TypeError(f"Unsupported image input type: {type(image)}")
+
+        tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".png")
+        pil.save(tmp, format="PNG")
+        tmp.close()
+        return tmp.name, True
+
+    def run(
+        self,
+        image: ImageInput,
+        prompt: str = "<image>\nFree OCR. ",
+        output_path: Optional[str] = None,
+        save_results: bool = False,
+        test_compress: bool = False,
+        **kwargs,
+    ) -> str:
+        """
+        Run OCR on an image and return the recognized text.
+
+        Args:
+            image: Path, bytes, or PIL Image.
+            prompt: Prompt passed to the vision model (keeps the default used
+                in the reference notebook).
+            output_path: Optional directory for saving debug outputs.
+            save_results: If True, Unsloth will save visualization artifacts.
+            test_compress: Forwarded to `model.infer`.
+            kwargs: Additional overrides for infer (base_size, image_size, etc).
+        """
+        self._load_model()
+
+        image_file, should_cleanup = self._prepare_image_file(image)
+        infer_kwargs = {
+            "prompt": prompt,
+            "image_file": image_file,
+            "output_path": output_path or "",
+            "base_size": kwargs.pop("base_size", self.base_size),
+            "image_size": kwargs.pop("image_size", self.image_size),
+            "crop_mode": kwargs.pop("crop_mode", self.crop_mode),
+            "save_results": save_results,
+            "test_compress": test_compress,
+        }
+
+        try:
+            result = self._model.infer(self._tokenizer, **infer_kwargs)
+        finally:
+            # Clean up temp files created from bytes/PIL inputs.
+            if should_cleanup and os.path.exists(image_file):
+                try:
+                    os.remove(image_file)
+                except OSError:
+                    pass
+
+        # Unsloth infer returns a dict-like object; stringify for callers.
+        if isinstance(result, str):
+            return result.strip()
+        if isinstance(result, dict) and "text" in result:
+            return str(result["text"]).strip()
+        return str(result).strip()
+
+
+def deepseek_ocr(
+    image: ImageInput,
+    prompt: str = "<image>\nFree OCR. ",
+    model_id: str = "unsloth/DeepSeek-OCR",
+    local_dir: Optional[str] = None,
+    **kwargs,
+) -> str:
+    """
+    Functional wrapper that mirrors the reference notebook defaults.
+
+    Example:
+        text = deepseek_ocr(\"invoice.png\")
+    """
+    runner = DeepSeekOCR(
+        model_id=model_id,
+        local_dir=local_dir or os.path.expanduser("~/.npcsh/models/deepseek_ocr"),
+        load_in_4bit=kwargs.pop("load_in_4bit", False),
+        base_size=kwargs.pop("base_size", 1024),
+        image_size=kwargs.pop("image_size", 640),
+        crop_mode=kwargs.pop("crop_mode", True),
+    )
+    return runner.run(
+        image=image,
+        prompt=prompt,
+        output_path=kwargs.pop("output_path", None),
+        save_results=kwargs.pop("save_results", False),
+        test_compress=kwargs.pop("test_compress", False),
+        **kwargs,
+    )

npcpy/memory/command_history.py

@@ -405,9 +405,13 @@ def save_kg_to_db(engine: Engine, kg_data: Dict[str, Any], team_name: str, npc_n
 def generate_message_id() -> str:
     return str(uuid.uuid4())
 
+
+
+from sqlalchemy import event, Table, Column, Integer, String, Text
+from sqlalchemy.orm import mapper
+
 class CommandHistory:
     def __init__(self, db: Union[str, Engine] = "~/npcsh_history.db"):
-        
         if isinstance(db, str):
             self.engine = create_engine_from_path(db)
             self.db_path = db
@@ -415,15 +419,54 @@ class CommandHistory:
             self.engine = db
             self.db_path = str(db.url)
         else:
-            raise TypeError(f"Unsupported type for CommandHistory db parameter: {type(db)}")
+            raise TypeError(f"Unsupported type: {type(db)}")
 
         self._initialize_schema()
-
+        self._setup_execution_triggers()
+        self.backfill_execution_tables()
+    def backfill_execution_tables(self):
+        with self.engine.begin() as conn:
+            conn.execute(text("""
+                INSERT OR IGNORE INTO jinx_executions 
+                (message_id, jinx_name, input, timestamp, npc, team, 
+                conversation_id)
+                SELECT 
+                    message_id,
+                    SUBSTR(content, 2, 
+                        CASE 
+                            WHEN INSTR(SUBSTR(content, 2), ' ') > 0 
+                            THEN INSTR(SUBSTR(content, 2), ' ') - 1
+                            ELSE LENGTH(content) - 1
+                        END
+                    ),
+                    content,
+                    timestamp,
+                    npc,
+                    team,
+                    conversation_id
+                FROM conversation_history
+                WHERE role = 'user' AND content LIKE '/%'
+            """))
+            
+            conn.execute(text("""
+                INSERT OR IGNORE INTO npc_executions 
+                (message_id, input, timestamp, npc, team, conversation_id, 
+                model, provider)
+                SELECT 
+                    message_id,
+                    content,
+                    timestamp,
+                    npc,
+                    team,
+                    conversation_id,
+                    model,
+                    provider
+                FROM conversation_history
+                WHERE role = 'user' AND npc IS NOT NULL
+            """))
     def _initialize_schema(self):
-        """Creates all necessary tables."""
         metadata = MetaData()
         
-        
         Table('command_history', metadata,
             Column('id', Integer, primary_key=True, autoincrement=True),
             Column('timestamp', String(50)),
@@ -433,7 +476,6 @@ class CommandHistory:
             Column('location', Text)
         )
         
-        
         Table('conversation_history', metadata,
             Column('id', Integer, primary_key=True, autoincrement=True),
             Column('message_id', String(50), unique=True, nullable=False),
@@ -448,33 +490,48 @@ class CommandHistory:
             Column('team', String(100))
         )
         
-        
         Table('message_attachments', metadata,
             Column('id', Integer, primary_key=True, autoincrement=True),
-            Column('message_id', String(50), ForeignKey('conversation_history.message_id', ondelete='CASCADE'), nullable=False),
+            Column('message_id', String(50), 
+                   ForeignKey('conversation_history.message_id', 
+                              ondelete='CASCADE'), 
+                   nullable=False),
             Column('attachment_name', String(255)),
             Column('attachment_type', String(100)),
             Column('attachment_data', LargeBinary),
             Column('attachment_size', Integer),
             Column('upload_timestamp', String(50)),
-            Column('file_path', Text) 
+            Column('file_path', Text)
+        )
+
+        Table('labels', metadata,
+            Column('id', Integer, primary_key=True, autoincrement=True),
+            Column('entity_type', String(50), nullable=False),
+            Column('entity_id', String(100), nullable=False),
+            Column('label', String(100), nullable=False),
+            Column('metadata', Text),
+            Column('created_at', DateTime, default=func.now())
+        )
+            
+        Table('jinx_executions', metadata,
+            Column('message_id', String(50), primary_key=True),
+            Column('jinx_name', String(100)),
+            Column('input', Text),
+            Column('timestamp', String(50)),
+            Column('npc', String(100)),
+            Column('team', String(100)),
+            Column('conversation_id', String(100))
         )
         
-        
-        Table('jinx_execution_log', metadata,
-            Column('execution_id', Integer, primary_key=True, autoincrement=True),
-            Column('triggering_message_id', String(50), ForeignKey('conversation_history.message_id', ondelete='CASCADE'), nullable=False),
-            Column('response_message_id', String(50), ForeignKey('conversation_history.message_id', ondelete='SET NULL')),
-            Column('conversation_id', String(100), nullable=False),
-            Column('timestamp', String(50), nullable=False),
-            Column('npc_name', String(100)),
-            Column('team_name', String(100)),
-            Column('jinx_name', String(100), nullable=False),
-            Column('jinx_inputs', Text),
-            Column('jinx_output', Text),
-            Column('status', String(50), nullable=False),
-            Column('error_message', Text),
-            Column('duration_ms', Integer)
+        Table('npc_executions', metadata,
+            Column('message_id', String(50), primary_key=True),
+            Column('input', Text),
+            Column('timestamp', String(50)),
+            Column('npc', String(100)),
+            Column('team', String(100)),
+            Column('conversation_id', String(100)),
+            Column('model', String(100)),
+            Column('provider', String(100))
         )
         
         Table('memory_lifecycle', metadata,
@@ -492,30 +549,137 @@ class CommandHistory:
             Column('provider', String(100)),
             Column('created_at', DateTime, default=func.now())
         )
-                
         
         metadata.create_all(self.engine, checkfirst=True)
+        init_kg_schema(self.engine)
+
+    def _setup_execution_triggers(self):
+        if 'sqlite' in str(self.engine.url):
+            with self.engine.begin() as conn:
+                conn.execute(text("""
+                    CREATE TRIGGER IF NOT EXISTS populate_jinx_executions
+                    AFTER INSERT ON conversation_history
+                    WHEN NEW.role = 'user' AND NEW.content LIKE '/%'
+                    BEGIN
+                        INSERT OR IGNORE INTO jinx_executions 
+                        (message_id, jinx_name, input, timestamp, npc, team, 
+                         conversation_id)
+                        VALUES (
+                            NEW.message_id,
+                            SUBSTR(NEW.content, 2, 
+                                CASE 
+                                    WHEN INSTR(SUBSTR(NEW.content, 2), ' ') > 0 
+                                    THEN INSTR(SUBSTR(NEW.content, 2), ' ') - 1
+                                    ELSE LENGTH(NEW.content) - 1
+                                END
+                            ),
+                            NEW.content,
+                            NEW.timestamp,
+                            NEW.npc,
+                            NEW.team,
+                            NEW.conversation_id
+                        );
+                    END
+                """))
+                
+                conn.execute(text("""
+                    CREATE TRIGGER IF NOT EXISTS populate_npc_executions
+                    AFTER INSERT ON conversation_history
+                    WHEN NEW.role = 'user' AND NEW.npc IS NOT NULL
+                    BEGIN
+                        INSERT OR IGNORE INTO npc_executions 
+                        (message_id, input, timestamp, npc, team, 
+                         conversation_id, model, provider)
+                        VALUES (
+                            NEW.message_id,
+                            NEW.content,
+                            NEW.timestamp,
+                            NEW.npc,
+                            NEW.team,
+                            NEW.conversation_id,
+                            NEW.model,
+                            NEW.provider
+                        );
+                    END
+                """))
+
+    def get_jinx_executions(self, jinx_name: str = None, limit: int = 1000) -> List[Dict]:
+        if jinx_name:
+            stmt = """
+                SELECT je.*, l.label
+                FROM jinx_executions je
+                LEFT JOIN labels l ON l.entity_type = 'message' 
+                    AND l.entity_id = je.message_id
+                WHERE je.jinx_name = :jinx_name
+                ORDER BY je.timestamp DESC
+                LIMIT :limit
+            """
+            return self._fetch_all(stmt, {"jinx_name": jinx_name, "limit": limit})
         
+        stmt = """
+            SELECT je.*, l.label
+            FROM jinx_executions je
+            LEFT JOIN labels l ON l.entity_type = 'message' 
+                AND l.entity_id = je.message_id
+            ORDER BY je.timestamp DESC
+            LIMIT :limit
+        """
+        return self._fetch_all(stmt, {"limit": limit})
+
+    def get_npc_executions(self, npc_name: str, limit: int = 1000) -> List[Dict]:
+        stmt = """
+            SELECT ne.*, l.label
+            FROM npc_executions ne
+            LEFT JOIN labels l ON l.entity_type = 'message' 
+                AND l.entity_id = ne.message_id
+            WHERE ne.npc = :npc_name
+            ORDER BY ne.timestamp DESC
+            LIMIT :limit
+        """
+        return self._fetch_all(stmt, {"npc_name": npc_name, "limit": limit})
+
+    def label_execution(self, message_id: str, label: str):
+        self.add_label('message', message_id, label)
         
+    def add_label(self, entity_type: str, entity_id: str, label: str, metadata: dict = None):
+        stmt = """
+            INSERT INTO labels (entity_type, entity_id, label, metadata)
+            VALUES (:entity_type, :entity_id, :label, :metadata)
+        """
         with self.engine.begin() as conn:
-            
-            index_queries = [
-                "CREATE INDEX IF NOT EXISTS idx_jinx_log_trigger_msg ON jinx_execution_log (triggering_message_id)",
-                "CREATE INDEX IF NOT EXISTS idx_jinx_log_convo_id ON jinx_execution_log (conversation_id)",
-                "CREATE INDEX IF NOT EXISTS idx_jinx_log_jinx_name ON jinx_execution_log (jinx_name)",
-                "CREATE INDEX IF NOT EXISTS idx_jinx_log_timestamp ON jinx_execution_log (timestamp)"
-            ]
-            
-            for idx_query in index_queries:
-                try:
-                    conn.execute(text(idx_query))
-                except SQLAlchemyError:
-                    
-                    pass
+            conn.execute(text(stmt), {
+                "entity_type": entity_type,
+                "entity_id": entity_id,
+                "label": label,
+                "metadata": json.dumps(metadata) if metadata else None
+            })
+
+    def get_labels(self, entity_type: str = None, label: str = None) -> List[Dict]:
+        conditions = []
+        params = {}
         
+        if entity_type:
+            conditions.append("entity_type = :entity_type")
+            params["entity_type"] = entity_type
+        if label:
+            conditions.append("label = :label")
+            params["label"] = label
         
-        init_kg_schema(self.engine)
+        where = f"WHERE {' AND '.join(conditions)}" if conditions else ""
+        stmt = f"SELECT * FROM labels {where} ORDER BY created_at DESC"
+        
+        return self._fetch_all(stmt, params)
 
+    def get_training_data_by_label(self, label: str = 'training') -> List[Dict]:
+        stmt = """
+            SELECT l.entity_type, l.entity_id, l.metadata,
+                ch.content, ch.role, ch.npc, ch.conversation_id
+            FROM labels l
+            LEFT JOIN conversation_history ch ON 
+                (l.entity_type = 'message' AND l.entity_id = ch.message_id)
+            WHERE l.label = :label
+        """
+        return self._fetch_all(stmt, {"label": label})
     def _execute_returning_id(self, stmt: str, params: Dict = None) -> Optional[int]:
         """Execute INSERT and return the generated ID"""
         with self.engine.begin() as conn:
@@ -535,6 +699,7 @@ class CommandHistory:
             result = conn.execute(text(stmt), params or {})
             return [dict(row._mapping) for row in result]
 
+
     def add_command(self, command, subcommands, output, location):
         timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
         stmt = """
@@ -1092,6 +1257,32 @@ def start_new_conversation(prepend: str = None) -> str:
         prepend = 'npcsh'
     return f"{prepend}_{datetime.now().strftime('%Y%m%d%H%M%S')}"
 
+
+def format_memory_context(memory_examples):
+    if not memory_examples:
+        return ""
+    
+    context_parts = []
+    
+    approved_examples = memory_examples.get("approved", [])
+    rejected_examples = memory_examples.get("rejected", [])
+    
+    if approved_examples:
+        context_parts.append("EXAMPLES OF GOOD MEMORIES:")
+        for ex in approved_examples[:5]:
+            final = ex.get("final_memory") or ex.get("initial_memory")
+            context_parts.append(f"- {final}")
+    
+    if rejected_examples:
+        context_parts.append("\nEXAMPLES OF POOR MEMORIES TO AVOID:")
+        for ex in rejected_examples[:3]:
+            context_parts.append(f"- {ex.get('initial_memory')}")
+    
+    if context_parts:
+        context_parts.append("\nLearn from these examples to generate similar high-quality memories.")
+        return "\n".join(context_parts)
+    
+    return ""
 def save_conversation_message(
     command_history: CommandHistory,
     conversation_id: str,