MHamdan's picture
Initial commit: SPARKNET framework
d520909
"""
Vision-Language Model Interface
Abstract interface for multimodal models that understand both
images and text. Used for document understanding, VQA, and
complex reasoning over visual content.
"""
from abc import abstractmethod
from dataclasses import dataclass, field
from enum import Enum
from typing import Any, Dict, List, Optional, Tuple, Union
from ..chunks.models import BoundingBox
from .base import (
BaseModel,
BatchableModel,
ImageInput,
ModelCapability,
ModelConfig,
)
class VLMTask(str, Enum):
"""Tasks that VLM models can perform."""
# Document understanding
DOCUMENT_QA = "document_qa"
DOCUMENT_SUMMARY = "document_summary"
DOCUMENT_CLASSIFICATION = "document_classification"
# Visual understanding
IMAGE_CAPTION = "image_caption"
IMAGE_QA = "image_qa"
VISUAL_GROUNDING = "visual_grounding"
# Extraction
FIELD_EXTRACTION = "field_extraction"
TABLE_UNDERSTANDING = "table_understanding"
CHART_UNDERSTANDING = "chart_understanding"
# Generation
OCR_CORRECTION = "ocr_correction"
TEXT_GENERATION = "text_generation"
# Other
GENERAL = "general"
@dataclass
class VLMConfig(ModelConfig):
"""Configuration for vision-language models."""
max_tokens: int = 2048
temperature: float = 0.1
top_p: float = 0.9
max_image_size: int = 1024 # Max dimension in pixels
image_detail: str = "high" # "low", "high", "auto"
system_prompt: Optional[str] = None
def __post_init__(self):
super().__post_init__()
if not self.name:
self.name = "vlm"
@dataclass
class VLMMessage:
"""A message in a VLM conversation."""
role: str # "system", "user", "assistant"
content: str
images: List[ImageInput] = field(default_factory=list)
image_regions: List[Optional[BoundingBox]] = field(default_factory=list)
@dataclass
class VLMResponse:
"""Response from a VLM model."""
text: str
confidence: float = 0.0
tokens_used: int = 0
finish_reason: str = "stop" # "stop", "length", "content_filter"
# Grounding information (if applicable)
grounded_regions: List[BoundingBox] = field(default_factory=list)
region_labels: List[str] = field(default_factory=list)
# Structured output (if requested)
structured_data: Optional[Dict[str, Any]] = None
# Processing info
processing_time_ms: float = 0.0
model_metadata: Dict[str, Any] = field(default_factory=dict)
@dataclass
class DocumentQAResult:
"""Result of document question answering."""
question: str
answer: str
confidence: float = 0.0
# Evidence grounding
evidence_regions: List[BoundingBox] = field(default_factory=list)
evidence_text: List[str] = field(default_factory=list)
page_references: List[int] = field(default_factory=list)
# Abstention
abstained: bool = False
abstention_reason: Optional[str] = None
@dataclass
class FieldExtractionVLMResult:
"""Result of field extraction using VLM."""
fields: Dict[str, Any] = field(default_factory=dict)
confidence_scores: Dict[str, float] = field(default_factory=dict)
# Grounding for each field
field_regions: Dict[str, BoundingBox] = field(default_factory=dict)
field_evidence: Dict[str, str] = field(default_factory=dict)
# Abstention tracking
abstained_fields: List[str] = field(default_factory=list)
abstention_reasons: Dict[str, str] = field(default_factory=dict)
overall_confidence: float = 0.0
class VisionLanguageModel(BatchableModel):
"""
Abstract base class for Vision-Language Models.
These models combine visual understanding with language
capabilities for tasks like document QA, field extraction,
and visual reasoning.
"""
def __init__(self, config: Optional[VLMConfig] = None):
super().__init__(config or VLMConfig(name="vlm"))
self.config: VLMConfig = self.config
def get_capabilities(self) -> List[ModelCapability]:
return [ModelCapability.VISION_LANGUAGE]
@abstractmethod
def generate(
self,
prompt: str,
images: List[ImageInput],
**kwargs
) -> VLMResponse:
"""
Generate a response given text prompt and images.
Args:
prompt: Text prompt/question
images: List of images for context
**kwargs: Additional generation parameters
Returns:
VLMResponse with generated text
"""
pass
def process_batch(
self,
inputs: List[Tuple[str, List[ImageInput]]],
**kwargs
) -> List[VLMResponse]:
"""
Process multiple prompt-image pairs.
Args:
inputs: List of (prompt, images) tuples
**kwargs: Additional parameters
Returns:
List of VLMResponses
"""
return [
self.generate(prompt, images, **kwargs)
for prompt, images in inputs
]
@abstractmethod
def chat(
self,
messages: List[VLMMessage],
**kwargs
) -> VLMResponse:
"""
Multi-turn conversation with images.
Args:
messages: Conversation history
**kwargs: Additional parameters
Returns:
VLMResponse for the conversation
"""
pass
def answer_question(
self,
question: str,
document_images: List[ImageInput],
context: Optional[str] = None,
**kwargs
) -> DocumentQAResult:
"""
Answer a question about document images.
Args:
question: Question to answer
document_images: Document page images
context: Optional additional context
**kwargs: Additional parameters
Returns:
DocumentQAResult with answer and evidence
"""
prompt = self._build_qa_prompt(question, context)
response = self.generate(prompt, document_images, **kwargs)
# Parse response for answer and confidence
answer, confidence, abstained, reason = self._parse_qa_response(response.text)
return DocumentQAResult(
question=question,
answer=answer,
confidence=confidence,
evidence_regions=response.grounded_regions,
abstained=abstained,
abstention_reason=reason
)
def extract_fields(
self,
images: List[ImageInput],
schema: Dict[str, Any],
**kwargs
) -> FieldExtractionVLMResult:
"""
Extract fields from document images according to a schema.
Args:
images: Document page images
schema: Field schema (JSON Schema or Pydantic-like)
**kwargs: Additional parameters
Returns:
FieldExtractionVLMResult with extracted values
"""
prompt = self._build_extraction_prompt(schema)
response = self.generate(prompt, images, **kwargs)
# Parse structured response
result = self._parse_extraction_response(response, schema)
return result
def summarize_document(
self,
images: List[ImageInput],
max_length: int = 500,
**kwargs
) -> str:
"""
Generate a summary of document images.
Args:
images: Document page images
max_length: Maximum summary length
**kwargs: Additional parameters
Returns:
Document summary text
"""
prompt = f"""Summarize this document in at most {max_length} characters.
Focus on the main points and key information.
Be concise and factual."""
response = self.generate(prompt, images, **kwargs)
return response.text
def classify_document(
self,
images: List[ImageInput],
categories: List[str],
**kwargs
) -> Tuple[str, float]:
"""
Classify document into predefined categories.
Args:
images: Document page images
categories: List of possible categories
**kwargs: Additional parameters
Returns:
Tuple of (category, confidence)
"""
categories_str = ", ".join(categories)
prompt = f"""Classify this document into one of these categories: {categories_str}
Respond with just the category name and confidence (0-1).
Format: CATEGORY: confidence
If you cannot confidently classify, respond with: UNKNOWN: 0.0"""
response = self.generate(prompt, images, **kwargs)
# Parse response
try:
parts = response.text.strip().split(":")
category = parts[0].strip().upper()
confidence = float(parts[1].strip()) if len(parts) > 1 else 0.5
# Validate category
category_upper = {c.upper(): c for c in categories}
if category in category_upper:
return category_upper[category], confidence
return "UNKNOWN", 0.0
except Exception:
return "UNKNOWN", 0.0
def _build_qa_prompt(
self,
question: str,
context: Optional[str] = None
) -> str:
"""Build prompt for document QA."""
prompt_parts = [
"You are analyzing a document image. Answer the following question based only on what you can see in the document.",
"",
"IMPORTANT RULES:",
"- Only use information visible in the document",
"- If the answer is not found, say 'NOT FOUND' and explain why",
"- Be precise and quote exact values when possible",
"- Indicate your confidence level (HIGH, MEDIUM, LOW)",
""
]
if context:
prompt_parts.extend([
"Additional context:",
context,
""
])
prompt_parts.extend([
f"Question: {question}",
"",
"Provide your answer in this format:",
"ANSWER: [your answer]",
"CONFIDENCE: [HIGH/MEDIUM/LOW]",
"EVIDENCE: [quote or describe where you found this information]"
])
return "\n".join(prompt_parts)
def _parse_qa_response(
self,
response_text: str
) -> Tuple[str, float, bool, Optional[str]]:
"""Parse QA response for answer, confidence, and abstention."""
lines = response_text.strip().split("\n")
answer = ""
confidence = 0.5
abstained = False
reason = None
for line in lines:
line_lower = line.lower()
if line_lower.startswith("answer:"):
answer = line.split(":", 1)[1].strip()
elif line_lower.startswith("confidence:"):
conf_str = line.split(":", 1)[1].strip().upper()
confidence = {"HIGH": 0.9, "MEDIUM": 0.6, "LOW": 0.3}.get(conf_str, 0.5)
# Check for abstention
if "not found" in answer.lower() or "cannot find" in answer.lower():
abstained = True
reason = answer
return answer, confidence, abstained, reason
def _build_extraction_prompt(self, schema: Dict[str, Any]) -> str:
"""Build prompt for field extraction."""
import json
schema_str = json.dumps(schema, indent=2)
prompt = f"""Extract the following fields from this document image.
SCHEMA:
{schema_str}
RULES:
- Only extract values that are clearly visible in the document
- For each field, provide the exact value and its location
- If a field is not found, mark it as null with confidence 0
- Be precise with numbers, dates, and proper nouns
Respond in valid JSON format matching the schema.
Include a "_confidence" object with confidence scores (0-1) for each field.
Include a "_evidence" object with the text snippet where each value was found.
"""
return prompt
def _parse_extraction_response(
self,
response: VLMResponse,
schema: Dict[str, Any]
) -> FieldExtractionVLMResult:
"""Parse extraction response into structured result."""
import json
result = FieldExtractionVLMResult()
try:
# Try to parse JSON from response
text = response.text.strip()
# Find JSON block if wrapped in markdown
if "```json" in text:
start = text.find("```json") + 7
end = text.find("```", start)
text = text[start:end].strip()
elif "```" in text:
start = text.find("```") + 3
end = text.find("```", start)
text = text[start:end].strip()
data = json.loads(text)
# Extract fields
for key, value in data.items():
if key.startswith("_"):
continue
result.fields[key] = value
# Extract confidence scores
if "_confidence" in data:
result.confidence_scores = data["_confidence"]
# Extract evidence
if "_evidence" in data:
result.field_evidence = data["_evidence"]
# Track abstentions
for field_name in schema.get("properties", {}).keys():
if field_name not in result.fields or result.fields[field_name] is None:
result.abstained_fields.append(field_name)
result.abstention_reasons[field_name] = "Field not found in document"
# Calculate overall confidence
if result.confidence_scores:
result.overall_confidence = sum(result.confidence_scores.values()) / len(result.confidence_scores)
except json.JSONDecodeError:
# Failed to parse - mark all fields as abstained
for field_name in schema.get("properties", {}).keys():
result.abstained_fields.append(field_name)
result.abstention_reasons[field_name] = "Failed to parse extraction response"
return result