remdb 0.3.242__py3-none-any.whl

rem/services/content/providers.py

@@ -0,0 +1,760 @@
+"""Content provider plugins for different file types."""
+
+import json
+import multiprocessing
+import random
+import subprocess
+import sys
+from abc import ABC, abstractmethod
+from pathlib import Path
+from typing import Any, Optional
+
+from loguru import logger
+
+from rem.utils.constants import (
+    AUDIO_CHUNK_TARGET_SECONDS,
+    AUDIO_CHUNK_WINDOW_SECONDS,
+    MIN_SILENCE_MS,
+    SILENCE_THRESHOLD_DB,
+    SUBPROCESS_TIMEOUT_SECONDS,
+    WAV_HEADER_MIN_BYTES,
+    WHISPER_COST_PER_MINUTE,
+)
+from rem.utils.files import temp_file_from_bytes
+from rem.utils.mime_types import get_extension
+
+
+class ContentProvider(ABC):
+    """Base class for content extraction providers."""
+
+    @property
+    @abstractmethod
+    def name(self) -> str:
+        """Provider name for logging/debugging."""
+        pass
+
+    @abstractmethod
+    def extract(self, content: bytes, metadata: dict[str, Any]) -> dict[str, Any]:
+        """
+        Extract text content from file bytes.
+
+        Args:
+            content: Raw file bytes
+            metadata: File metadata (size, type, etc.)
+
+        Returns:
+            dict with:
+                - text: Extracted text content
+                - metadata: Additional metadata from extraction (optional)
+        """
+        pass
+
+
+class TextProvider(ContentProvider):
+    """
+    Text content provider for plain text formats.
+
+    Supports:
+    - Markdown (.md, .markdown) - With heading detection
+    - JSON (.json) - Pretty-printed text extraction
+    - YAML (.yaml, .yml) - Text extraction
+    - Plain text (.txt) - Direct UTF-8 extraction
+    - Code files (.py, .js, .ts, etc.) - Source code as text
+
+    Simple UTF-8 text extraction with basic metadata.
+    Future: Could add frontmatter parsing, JSON schema validation, etc.
+    """
+
+    @property
+    def name(self) -> str:
+        return "text"
+
+    def extract(self, content: bytes, metadata: dict[str, Any]) -> dict[str, Any]:
+        """
+        Extract text content from plain text files.
+
+        Args:
+            content: Text file bytes
+            metadata: File metadata
+
+        Returns:
+            dict with text and optional metadata (line count, headings for markdown, etc.)
+        """
+        # Decode UTF-8 (with fallback to latin-1)
+        try:
+            text = content.decode("utf-8")
+        except UnicodeDecodeError:
+            logger.debug("UTF-8 decode failed, falling back to latin-1")
+            text = content.decode("latin-1")
+
+        # Basic text analysis
+        lines = text.split("\n")
+
+        # Detect headings (for markdown files)
+        headings = [line for line in lines if line.strip().startswith("#")]
+
+        extraction_metadata = {
+            "line_count": len(lines),
+            "heading_count": len(headings) if headings else None,
+            "char_count": len(text),
+            "encoding": "utf-8",
+        }
+
+        return {
+            "text": text,
+            "metadata": extraction_metadata,
+        }
+
+
+class DocProvider(ContentProvider):
+    """
+    Document content provider using Kreuzberg.
+
+    Supports multiple document formats via Kreuzberg:
+    - PDF (.pdf) - Text extraction with OCR fallback
+    - Word (.docx) - Native format support
+    - PowerPoint (.pptx) - Slide content extraction
+    - Excel (.xlsx) - Spreadsheet data extraction
+    - Images (.png, .jpg) - OCR text extraction
+
+    Handles:
+    - Text extraction with OCR fallback
+    - Table detection and extraction
+    - Daemon process workaround for multiprocessing restrictions
+    """
+
+    @property
+    def name(self) -> str:
+        return "doc"
+
+    def _is_daemon_process(self) -> bool:
+        """Check if running in a daemon process."""
+        try:
+            return multiprocessing.current_process().daemon
+        except Exception:
+            return False
+
+    def _parse_in_subprocess(self, file_path: Path) -> dict:
+        """Run kreuzberg in a separate subprocess to bypass daemon restrictions."""
+        script = """
+import json
+import sys
+from pathlib import Path
+from kreuzberg import ExtractionConfig, extract_file_sync
+
+# Parse document with kreuzberg 3.x
+config = ExtractionConfig(
+    extract_tables=True,
+    chunk_content=False,
+    extract_keywords=False,
+)
+
+result = extract_file_sync(Path(sys.argv[1]), config=config)
+
+# Serialize result to JSON
+output = {
+    'content': result.content,
+    'tables': [t.model_dump() for t in result.tables] if result.tables else [],
+    'metadata': result.metadata
+}
+print(json.dumps(output))
+"""
+
+        # Run in subprocess
+        result = subprocess.run(
+            [sys.executable, "-c", script, str(file_path)],
+            capture_output=True,
+            text=True,
+            timeout=SUBPROCESS_TIMEOUT_SECONDS,
+        )
+
+        if result.returncode != 0:
+            raise RuntimeError(f"Subprocess parsing failed: {result.stderr}")
+
+        return json.loads(result.stdout)
+
+    def extract(self, content: bytes, metadata: dict[str, Any]) -> dict[str, Any]:
+        """
+        Extract document content using Kreuzberg.
+
+        Args:
+            content: Document file bytes
+            metadata: File metadata (should include content_type or extension)
+
+        Returns:
+            dict with text and extraction metadata
+        """
+        # Write bytes to temp file for kreuzberg
+        # Detect extension from metadata
+        content_type = metadata.get("content_type", "")
+        suffix = get_extension(content_type, default=".pdf")
+
+        with temp_file_from_bytes(content, suffix=suffix) as tmp_path:
+            # Check if running in daemon process
+            if self._is_daemon_process():
+                logger.info("Daemon process detected - using subprocess workaround for document parsing")
+                try:
+                    result_dict = self._parse_in_subprocess(tmp_path)
+                    text = result_dict["content"]
+                    extraction_metadata = {
+                        "table_count": len(result_dict["tables"]),
+                        "parser": "kreuzberg_subprocess",
+                        "file_extension": tmp_path.suffix,
+                    }
+                except Exception as e:
+                    logger.error(f"Subprocess parsing failed: {e}. Falling back to text-only.")
+                    # Fallback to simple text extraction (kreuzberg 3.x API)
+                    from kreuzberg import ExtractionConfig, extract_file_sync
+                    config = ExtractionConfig(extract_tables=False)
+                    result = extract_file_sync(tmp_path, config=config)
+                    text = result.content
+                    extraction_metadata = {
+                        "parser": "kreuzberg_fallback",
+                        "file_extension": tmp_path.suffix,
+                    }
+            else:
+                # Normal execution (not in daemon) - kreuzberg 4.x with native ONNX/Rust
+                from kreuzberg import ExtractionConfig, extract_file_sync
+                config = ExtractionConfig(
+                    enable_quality_processing=True,  # Enables table extraction with native ONNX
+                    chunk_content=False,  # We handle chunking ourselves
+                    extract_tables=False,  # Disable table extraction to avoid PyTorch dependency
+                )
+                result = extract_file_sync(tmp_path, config=config)
+                text = result.content
+                extraction_metadata = {
+                    "table_count": len(result.tables) if result.tables else 0,
+                    "parser": "kreuzberg",
+                    "file_extension": tmp_path.suffix,
+                }
+
+            return {
+                "text": text,
+                "metadata": extraction_metadata,
+            }
+
+
+class AudioProvider(ContentProvider):
+    """
+    Audio content provider using AudioChunker + OpenAI Whisper.
+
+    Handles:
+    - Audio chunking by silence near minute boundaries
+    - Transcription via OpenAI Whisper API
+    - Converts chunks to markdown format
+    - Supports WAV, M4A, MP3, FLAC, OGG (via pydub + ffmpeg)
+
+    Process:
+    1. Write audio bytes to temp file
+    2. Chunk audio by silence (AudioChunker)
+    3. Transcribe chunks (AudioTranscriber)
+    4. Combine into markdown format with timestamps
+    5. Clean up temp files
+
+    Returns markdown-formatted transcription that integrates
+    seamlessly with ContentService's markdown → chunk → embed pipeline.
+    """
+
+    @property
+    def name(self) -> str:
+        return "audio"
+
+    def extract(self, content: bytes, metadata: dict[str, Any]) -> dict[str, Any]:
+        """
+        Extract audio content via transcription.
+
+        Args:
+            content: Audio file bytes
+            metadata: File metadata (size, type, etc.)
+
+        Returns:
+            dict with:
+                - text: Markdown-formatted transcription with timestamps
+                - metadata: Extraction metadata (chunk_count, duration, cost)
+
+        Raises:
+            RuntimeError: If transcription fails or pydub not available
+            ValueError: If OpenAI API key missing
+        """
+        # Handle empty or invalid content
+        if not content or len(content) < WAV_HEADER_MIN_BYTES:
+            logger.warning("Audio content too small to be valid WAV file")
+            return {
+                "text": "[Invalid or empty audio file]",
+                "metadata": {"error": "invalid_content", "size": len(content)},
+            }
+
+        # Check for OpenAI API key (use settings)
+        from rem.settings import settings
+        api_key = settings.llm.openai_api_key
+        if not api_key:
+            logger.warning("No OpenAI API key found - audio transcription disabled")
+            return {
+                "text": "[Audio transcription requires LLM__OPENAI_API_KEY to be set]",
+                "metadata": {"error": "missing_api_key"},
+            }
+
+        # Import audio services (lazy import)
+        try:
+            from rem.services.audio import AudioChunker, AudioTranscriber
+        except ImportError as e:
+            logger.error(f"Audio services not available: {e}")
+            return {
+                "text": "[Audio processing requires: pip install rem[audio]]",
+                "metadata": {"error": "missing_dependencies"},
+            }
+
+        # Write bytes to temp file
+        # Detect extension from metadata or use .wav as fallback
+        content_type = metadata.get("content_type", "audio/wav")
+        extension = get_extension(content_type, default=".wav")
+
+        chunker = None
+        chunks = None
+
+        with temp_file_from_bytes(content, suffix=extension) as tmp_path:
+            try:
+                logger.info(f"Processing audio file: {tmp_path.name} ({len(content) / 1024 / 1024:.1f} MB)")
+
+                # Step 1: Chunk audio by silence
+                chunker = AudioChunker(
+                    target_chunk_seconds=AUDIO_CHUNK_TARGET_SECONDS,
+                    chunk_window_seconds=AUDIO_CHUNK_WINDOW_SECONDS,
+                    silence_threshold_db=SILENCE_THRESHOLD_DB,
+                    min_silence_ms=MIN_SILENCE_MS,
+                )
+
+                chunks = chunker.chunk_audio(tmp_path)
+                logger.info(f"Created {len(chunks)} audio chunks")
+
+                # Step 2: Transcribe chunks
+                transcriber = AudioTranscriber(api_key=api_key)
+                results = transcriber.transcribe_chunks(chunks)
+                logger.info(f"Transcribed {len(results)} chunks")
+
+                # Step 3: Combine into markdown format
+                # Format: Each chunk becomes a section with timestamp
+                markdown_parts = []
+                for result in results:
+                    timestamp = f"{result.start_seconds:.1f}s - {result.end_seconds:.1f}s"
+                    markdown_parts.append(f"## [{timestamp}]\n\n{result.text}\n")
+
+                markdown_text = "\n".join(markdown_parts)
+
+                # Calculate metadata
+                total_duration = sum(r.duration_seconds for r in results)
+                estimated_cost = (total_duration / 60) * WHISPER_COST_PER_MINUTE
+                successful_chunks = sum(1 for r in results if r.confidence > 0)
+
+                extraction_metadata = {
+                    "chunk_count": len(chunks),
+                    "transcribed_chunks": successful_chunks,
+                    "duration_seconds": total_duration,
+                    "estimated_cost": estimated_cost,
+                    "parser": "whisper_api",
+                }
+
+                logger.info(
+                    f"Transcription complete: {successful_chunks}/{len(chunks)} chunks, "
+                    f"${estimated_cost:.3f} cost"
+                )
+
+                return {
+                    "text": markdown_text,
+                    "metadata": extraction_metadata,
+                }
+
+            except Exception as e:
+                logger.error(f"Audio extraction failed: {e}")
+                raise RuntimeError(f"Audio transcription failed: {e}") from e
+
+            finally:
+                # Clean up audio chunks (temp file cleanup handled by context manager)
+                if chunker is not None and chunks is not None:
+                    try:
+                        chunker.cleanup_chunks(chunks)
+                    except Exception as e:
+                        logger.warning(f"Chunk cleanup failed: {e}")
+
+
+class SchemaProvider(ContentProvider):
+    """
+    Schema content provider for agent/evaluator schemas.
+
+    Detects and processes YAML/JSON files containing:
+    - Agent schemas (type: object with json_schema_extra.kind: agent and json_schema_extra.name: <name>)
+    - Evaluator schemas (type: object with json_schema_extra.kind: evaluator and json_schema_extra.name: <name>)
+
+    Stores schemas in the schemas table with deterministic IDs for upsert by name.
+
+    Pattern:
+    - Checks for schema markers (type: object + kind + name)
+    - Generates deterministic ID for upsert (tenant+schema_name)
+    - Stores full schema JSON in schemas table
+    - Extracts metadata (version, tags, provider_configs, embedding_fields)
+    """
+
+    @property
+    def name(self) -> str:
+        return "schema"
+
+    def extract(self, content: bytes, metadata: dict[str, Any]) -> dict[str, Any]:
+        """
+        Extract and validate agent/evaluator schema.
+
+        Args:
+            content: YAML or JSON file bytes
+            metadata: File metadata
+
+        Returns:
+            dict with:
+                - text: Human-readable schema summary
+                - metadata: Schema metadata
+                - schema_data: Full schema dict for storage
+                - is_schema: True if valid schema detected
+
+        Raises:
+            ValueError: If schema is invalid
+        """
+        import json
+        import yaml
+        from uuid import uuid5, NAMESPACE_DNS
+
+        # Decode content
+        try:
+            text = content.decode("utf-8")
+        except UnicodeDecodeError:
+            text = content.decode("latin-1")
+
+        # Try to parse as YAML/JSON
+        if metadata.get("content_type") == "application/json":
+            schema_data = json.loads(text)  # Raises JSONDecodeError on invalid JSON
+        else:
+            # Try YAML first (supports both YAML and JSON)
+            schema_data = yaml.safe_load(text)  # Raises yaml.YAMLError on invalid YAML
+
+        # Check if it's a schema (type: object + json_schema_extra.kind + json_schema_extra.name)
+        if not isinstance(schema_data, dict):
+            return {
+                "text": text,
+                "metadata": {"parser": "schema_fallback"},
+                "is_schema": False,
+            }
+
+        # Check for schema markers
+        is_object_type = schema_data.get("type") == "object"
+        json_schema_extra = schema_data.get("json_schema_extra", {})
+        kind = json_schema_extra.get("kind", "")
+        schema_name = json_schema_extra.get("name", "")
+
+        # Must have type: object, kind (agent or evaluator), and name
+        is_agent_schema = is_object_type and kind == "agent" and schema_name
+        is_evaluator_schema = is_object_type and kind == "evaluator" and schema_name
+
+        if not (is_agent_schema or is_evaluator_schema):
+            return {
+                "text": text,
+                "metadata": {"parser": "schema_fallback"},
+                "is_schema": False,
+            }
+
+        # Extract schema metadata
+        schema_type = kind  # "agent" or "evaluator"
+        version = json_schema_extra.get("version", "1.0.0")
+        tags = json_schema_extra.get("tags", [])
+
+        # Use name directly (already in kebab-case format)
+        short_name = schema_name
+
+        # Build human-readable summary
+        description = schema_data.get("description", "No description provided")
+        description_preview = description[:200] + "..." if len(description) > 200 else description
+
+        properties = schema_data.get("properties", {})
+        required_fields = schema_data.get("required", [])
+
+        summary_parts = [
+            f"# {schema_type.title()} Schema: {short_name}",
+            f"**Version:** {version}",
+            f"**Name:** {schema_name}",
+            f"**Kind:** {kind}",
+            "",
+            "## Description",
+            description_preview,
+            "",
+            "## Output Fields",
+        ]
+
+        for field_name, field_spec in list(properties.items())[:10]:  # Limit to 10 fields
+            field_type = field_spec.get("type", "unknown")
+            field_desc = field_spec.get("description", "")
+            required = " (required)" if field_name in required_fields else ""
+            summary_parts.append(f"- **{field_name}**: {field_type}{required} - {field_desc[:50]}")
+
+        if len(properties) > 10:
+            summary_parts.append(f"- ... and {len(properties) - 10} more fields")
+
+        text_summary = "\n".join(summary_parts)
+
+        # Extract additional metadata
+        extraction_metadata = {
+            "parser": "schema",
+            "schema_type": schema_type,
+            "short_name": short_name,
+            "version": version,
+            "kind": kind,
+            "name": schema_name,
+            "tags": tags,
+            "field_count": len(properties),
+            "required_field_count": len(required_fields),
+            "provider_configs": json_schema_extra.get("provider_configs", []),
+            "embedding_fields": json_schema_extra.get("embedding_fields", []),
+            "category": json_schema_extra.get("category"),
+        }
+
+        return {
+            "text": text_summary,
+            "metadata": extraction_metadata,
+            "schema_data": schema_data,
+            "is_schema": True,
+        }
+
+
+class ImageProvider(ContentProvider):
+    """
+    Image content provider with vision LLM analysis and CLIP embeddings.
+
+    Features:
+    - Tier-based vision analysis (gold tier always gets analysis)
+    - Sampling-based vision analysis for non-gold users
+    - Vision LLM description generation (Anthropic, Gemini, OpenAI)
+    - Future: CLIP embeddings for semantic image search
+
+    Process:
+    1. Check user tier and sampling rate
+    2. If eligible, run vision LLM analysis
+    3. Extract image metadata (dimensions, format)
+    4. Return markdown description or basic metadata
+    5. Save to ImageResource table (not Resource)
+
+    Vision analysis is expensive, so it's gated by:
+    - User tier (gold = always, silver/free = sampled)
+    - Sample rate setting (0.0 = never, 1.0 = always)
+    """
+
+    def __init__(self, user_tier: Optional[str] = None):
+        """
+        Initialize image provider.
+
+        Args:
+            user_tier: User tier (free, silver, gold) for vision gating
+        """
+        self.user_tier = user_tier
+
+    @property
+    def name(self) -> str:
+        return "image"
+
+    def _should_analyze_with_vision(self, sample_rate: float) -> bool:
+        """
+        Determine if image should get vision LLM analysis.
+
+        Args:
+            sample_rate: Sampling rate from settings (0.0-1.0)
+
+        Returns:
+            True if should analyze, False otherwise
+        """
+        # Import here to avoid circular dependency
+        from rem.models.entities import UserTier
+
+        # Gold tier always gets vision analysis
+        if self.user_tier == UserTier.GOLD.value:
+            logger.info("Gold tier user - vision analysis enabled")
+            return True
+
+        # For non-gold users, use sampling
+        if sample_rate > 0.0:
+            should_analyze = random.random() < sample_rate
+            if should_analyze:
+                logger.info(f"Vision analysis sampled (rate={sample_rate})")
+            return should_analyze
+
+        return False
+
+    def extract(self, content: bytes, metadata: dict[str, Any]) -> dict[str, Any]:
+        """
+        Extract image content with optional vision LLM analysis.
+
+        Args:
+            content: Image file bytes
+            metadata: File metadata (size, type, etc.)
+
+        Returns:
+            dict with:
+                - text: Markdown description (if vision enabled) or basic metadata
+                - metadata: Extraction metadata (dimensions, format, vision info)
+                - image_specific: Additional image metadata for ImageResource
+
+        Raises:
+            RuntimeError: If vision analysis fails
+        """
+        # Import settings here to avoid circular dependency
+        from rem.settings import settings
+
+        # Extract basic image metadata using PIL
+        try:
+            from PIL import Image
+            import io
+
+            img = Image.open(io.BytesIO(content))
+            image_width = img.width
+            image_height = img.height
+            image_format = img.format or "UNKNOWN"
+        except ImportError:
+            logger.warning("PIL not available - image metadata extraction disabled")
+            image_width = None
+            image_height = None
+            image_format = None
+        except Exception as e:
+            logger.warning(f"Failed to extract image metadata: {e}")
+            image_width = None
+            image_height = None
+            image_format = None
+
+        # Check if vision analysis should be performed
+        sample_rate = settings.content.image_vllm_sample_rate
+        should_analyze = self._should_analyze_with_vision(sample_rate)
+
+        vision_description = None
+        vision_provider = None
+        vision_model = None
+
+        if should_analyze:
+            # Perform vision LLM analysis
+            try:
+                from rem.utils.vision import ImageAnalyzer, VisionProvider
+
+                # Get provider from settings
+                provider_str = settings.content.image_vllm_provider.lower()
+                provider_map = {
+                    "anthropic": VisionProvider.ANTHROPIC,
+                    "gemini": VisionProvider.GEMINI,
+                    "openai": VisionProvider.OPENAI,
+                }
+                provider = provider_map.get(provider_str, VisionProvider.ANTHROPIC)
+
+                # Create analyzer
+                analyzer = ImageAnalyzer(
+                    provider=provider,
+                    model=settings.content.image_vllm_model,
+                )
+
+                # Write bytes to temp file for analysis
+                content_type = metadata.get("content_type", "image/png")
+                extension = get_extension(content_type, default=".png")
+
+                with temp_file_from_bytes(content, suffix=extension) as tmp_path:
+                    # Analyze image
+                    result = analyzer.analyze_image(tmp_path)
+                    vision_description = result.description
+                    vision_provider = result.provider.value
+                    vision_model = result.model
+
+                    logger.info(f"Vision analysis complete: {len(vision_description)} chars")
+
+            except ImportError as e:
+                logger.warning(f"Vision analysis not available: {e}")
+            except Exception as e:
+                logger.error(f"Vision analysis failed: {e}")
+
+        # Build text content
+        if vision_description:
+            # Use vision description as primary content
+            text = f"# Image Analysis\n\n{vision_description}"
+            if image_width and image_height:
+                text += f"\n\n**Image Details:** {image_width}x{image_height} {image_format}"
+        else:
+            # Fallback to basic metadata
+            if image_width and image_height:
+                text = f"**Image:** {image_width}x{image_height} {image_format}"
+            else:
+                text = "**Image:** Metadata extraction unavailable"
+
+        # Generate CLIP embedding (if Jina API key available)
+        clip_embedding = None
+        clip_dimensions = None
+        clip_tokens = None
+
+        try:
+            from rem.utils.clip_embeddings import JinaCLIPEmbedder
+
+            # Only attempt CLIP embeddings if using Jina provider
+            if settings.content.clip_provider != "jina":
+                logger.debug(
+                    f"CLIP provider set to '{settings.content.clip_provider}' - "
+                    "skipping Jina embeddings (self-hosted not yet implemented)"
+                )
+            else:
+                embedder = JinaCLIPEmbedder(
+                    api_key=settings.content.jina_api_key,
+                    model=settings.content.clip_model,
+                )
+
+                if embedder.is_available():
+                    # Write bytes to temp file for CLIP embedding
+                    content_type = metadata.get("content_type", "image/png")
+                    extension = get_extension(content_type, default=".png")
+
+                    with temp_file_from_bytes(content, suffix=extension) as tmp_path:
+                        # Generate CLIP embedding
+                        result = embedder.embed_image(tmp_path)
+                        if result:
+                            clip_embedding = result.embedding  # type: ignore[attr-defined]
+                            clip_dimensions = result.dimensions  # type: ignore[attr-defined]
+                            clip_tokens = result.tokens_used  # type: ignore[attr-defined]
+                            logger.info(
+                                f"CLIP embedding generated: {clip_dimensions} dims, {clip_tokens} tokens"
+                            )
+                else:
+                    logger.debug(
+                        "CLIP embeddings disabled - set CONTENT__JINA_API_KEY to enable. "
+                        "Get free API key at https://jina.ai/embeddings/"
+                    )
+
+        except ImportError:
+            logger.debug("CLIP embedding module not available")
+        except Exception as e:
+            logger.warning(f"CLIP embedding generation failed (non-fatal): {e}")
+
+        # Build extraction metadata
+        extraction_metadata = {
+            "parser": "image_provider",
+            "vision_enabled": vision_description is not None,
+            "vision_provider": vision_provider,
+            "vision_model": vision_model,
+            "image_width": image_width,
+            "image_height": image_height,
+            "image_format": image_format,
+            "clip_enabled": clip_embedding is not None,
+            "clip_dimensions": clip_dimensions,
+            "clip_tokens": clip_tokens,
+        }
+
+        # Add image-specific metadata for ImageResource
+        image_specific = {
+            "image_width": image_width,
+            "image_height": image_height,
+            "image_format": image_format,
+            "vision_description": vision_description,
+            "vision_provider": vision_provider,
+            "vision_model": vision_model,
+            "clip_embedding": clip_embedding,
+            "clip_dimensions": clip_dimensions,
+        }
+
+        return {
+            "text": text,
+            "metadata": extraction_metadata,
+            "image_specific": image_specific,
+        }