diff options
Diffstat (limited to 'src/server/chunker/pdf_chunker.py')
| -rw-r--r-- | src/server/chunker/pdf_chunker.py | 187 |
1 files changed, 71 insertions, 116 deletions
diff --git a/src/server/chunker/pdf_chunker.py b/src/server/chunker/pdf_chunker.py index 697550f2e..e34753176 100644 --- a/src/server/chunker/pdf_chunker.py +++ b/src/server/chunker/pdf_chunker.py @@ -153,7 +153,7 @@ class ElementExtractor: xref = img_info[0] # XREF of the image in the PDF base_image = page.parent.extract_image(xref) # Extract the image by its XREF image_bytes = base_image["image"] - image = Image.open(io.BytesIO(image_bytes)) # Convert bytes to PIL image + image = Image.open(io.BytesIO(image_bytes)).convert("RGB") # Ensure it's RGB before saving as PNG width_ratio = img.width / page.rect.width # Scale factor for width height_ratio = img.height / page.rect.height # Scale factor for height @@ -276,12 +276,13 @@ class PDFChunker: :param output_folder: Folder to store the output files (extracted tables/images). :param image_batch_size: The batch size for processing visual elements. """ - self.client = Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY")) # Initialize the Anthropic API client + self.client = OpenAI() # ← replaces Anthropic() self.output_folder = output_folder self.image_batch_size = image_batch_size # Batch size for image processing self.doc_id = doc_id # Add doc_id self.element_extractor = ElementExtractor(output_folder, doc_id) + async def chunk_pdf(self, file_data: bytes, file_name: str, doc_id: str, job_id: str) -> List[Dict[str, Any]]: """ Processes a PDF file, extracting text and visual elements, and returning structured chunks. @@ -518,124 +519,77 @@ class PDFChunker: def batch_summarize_images(self, images: Dict[int, str]) -> Dict[int, str]: """ - Summarize images or tables by generating descriptive text. - - :param images: A dictionary mapping image numbers to base64-encoded image data. - :return: A dictionary mapping image numbers to their generated summaries. - """ - # Prompt for the AI model to summarize images and tables - prompt = f"""<instruction> - <task> - You are tasked with summarizing a series of {len(images)} images and tables for use in a RAG (Retrieval-Augmented Generation) system. - Your goal is to create concise, informative summaries that capture the essential content of each image or table. - These summaries will be used for embedding, so they should be descriptive and relevant. The image or table will be outlined in red on an image of the full page that it is on. Where necessary, use the context of the full page to heklp with the summary but don't summarize other content on the page. - </task> - - <steps> - <step>Identify whether it's an image or a table.</step> - <step>Examine its content carefully.</step> - <step> - Write a detailed summary that captures the main points or visual elements: - <details> - <table>After summarizing what the table is about, include the column headers, a detailed summary of the data, and any notable data trends.</table> - <image>Describe the main subjects, actions, or notable features.</image> - </details> - </step> - <step>Focus on writing summaries that would make it easy to retrieve the content if compared to a user query using vector similarity search.</step> - <step>Keep summaries concise and include important words that may help with retrieval (but do not include numbers and numerical data).</step> - </steps> - - <important_notes> - <note>Avoid using special characters like &, <, >, ", ', $, %, etc. Instead, use their word equivalents:</note> - <note>Use "and" instead of &.</note> - <note>Use "dollars" instead of $.</note> - <note>Use "percent" instead of %.</note> - <note>Refrain from using quotation marks " or apostrophes ' unless absolutely necessary.</note> - <note>Ensure your output is in valid XML format.</note> - </important_notes> - - <formatting> - <note>Enclose all summaries within a root element called <summaries>.</note> - <note>Use <summary> tags to enclose each individual summary.</note> - <note>Include an attribute 'number' in each <summary> tag to indicate the sequence, matching the provided image numbers.</note> - <note>Start each summary by indicating whether it's an image or a table (e.g., "This image shows..." or "The table presents...").</note> - <note>If an image is completely blank, leave the summary blank (e.g., <summary number="3"></summary>).</note> - </formatting> - - <example> - <note>Do not replicate the example below—stay grounded to the content of the table or image and describe it completely and accurately.</note> - <output> - <summaries> - <summary number="1"> - The image shows two men shaking hands on stage at a formal event. The man on the left, in a dark suit and glasses, has a professional appearance, possibly an academic or business figure. The man on the right, Tim Cook, CEO of Apple, is recognizable by his silver hair and dark blue blazer. Cook holds a document titled "Tsinghua SEM EMBA," suggesting a link to Tsinghua University’s Executive MBA program. The backdrop displays English and Chinese text about business management and education, with the event dated October 23, 2014. - </summary> - <summary number="2"> - The table compares the company's assets between December 30, 2023, and September 30, 2023. Key changes include an increase in cash and cash equivalents, while marketable securities had a slight rise. Accounts receivable and vendor non-trade receivables decreased. Inventories and other current assets saw minor fluctuations. Non-current assets like marketable securities slightly declined, while property, plant, and equipment remained stable. Total assets showed minimal change, holding steady at around three hundred fifty-three billion dollars. - </summary> - <summary number="3"> - The table outlines the company's shareholders' equity as of December 30, 2023, versus September 30, 2023. Common stock and additional paid-in capital increased, and retained earnings shifted from a deficit to a positive figure. Accumulated other comprehensive loss decreased. Overall, total shareholders' equity rose significantly, while total liabilities and equity remained nearly unchanged at about three hundred fifty-three billion dollars. - </summary> - <summary number="4"> - The table details the company's liabilities as of December 30, 2023, compared to September 30, 2023. Current liabilities decreased due to lower accounts payable and other current liabilities, while deferred revenue slightly increased. Commercial paper significantly decreased, and term debt rose modestly. Non-current liabilities were stable, with minimal changes in term debt and other non-current liabilities. Total liabilities dropped from two hundred ninety billion dollars to two hundred seventy-nine billion dollars. - </summary> - <summary number="5"> - </summary> - </summaries> - </output> - </example> - - <final_notes> - <note>Process each image or table in the order provided.</note> - <note>Maintain consistent formatting throughout your response.</note> - <note>Ensure the output is in full, valid XML format with the root <summaries> element and each summary being within a <summary> element with the summary number specified as well.</note> - </final_notes> -</instruction> - """ - content = [] - for number, img in images.items(): - content.append({"type": "text", "text": f"\nImage {number}:\n"}) - content.append({"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": img}}) + Summarise a batch of images/tables with GPT‑4o using Structured Outputs. + :param images: {image_number: base64_png} + :return: {image_number: summary_text} + """ + # -------- 1. Build the prompt ----------- + content: list[dict] = [] + for n, b64 in images.items(): + content.append({"type": "text", + "text": f"\nImage {n} (outlined in red on the page):"}) + content.append({"type": "image_url", + "image_url": {"url": f"data:image/png;base64,{b64}"}}) messages = [ - {"role": "user", "content": content} + { + "role": "system", + "content": ( + "You are generating retrieval‑ready summaries for each highlighted " + "image or table. Start by identifying whether the element is an " + "image or a table, then write one informative sentence that a vector " + "search would find useful. Provide detail but limit to a couple of paragraphs per image." + ), + }, + {"role": "user", "content": content}, ] + schema = { + "type": "object", + "properties": { + "summaries": { + "type": "array", + "items": { + "type": "object", + "properties": { + "number": {"type": "integer"}, + "type": {"type": "string", "enum": ["image", "table"]}, + "summary": {"type": "string"} + }, + "required": ["number", "type", "summary"], + "additionalProperties": False + } + } + }, + "required": ["summaries"], + "additionalProperties": False + } + + # ---------- OpenAI call ----------------------------------------------------- try: - response = self.client.messages.create( - model='claude-3-5-sonnet-20240620', - system=prompt, - max_tokens=400 * len(images), # Increased token limit for more detailed summaries + resp = self.client.chat.completions.create( + model="gpt-4o", messages=messages, + max_tokens=400 * len(images), temperature=0, - extra_headers={"anthropic-beta": "max-tokens-3-5-sonnet-2024-07-15"} + response_format={ + "type": "json_schema", + "json_schema": { + "name": "image_batch_summaries", # ← REQUIRED + "schema": schema, # ← REQUIRED + "strict": True # ← strongly recommended + }, + }, ) - # Parse the response - text = response.content[0].text - #print(text) - # Attempt to parse and fix the XML if necessary - parser = etree.XMLParser(recover=True) - root = etree.fromstring(text, parser=parser) - # Check if there were errors corrected - # if parser.error_log: - # #print("XML Parsing Errors:") - # for error in parser.error_log: - # #print(error) - # Extract summaries - summaries = {} - for summary in root.findall('summary'): - number = int(summary.get('number')) - content = summary.text.strip() if summary.text else "" - if content: # Only include non-empty summaries - summaries[number] = content - - return summaries + parsed = json.loads(resp.choices[0].message.content) # schema‑safe + return {item["number"]: item["summary"] + for item in parsed["summaries"]} except Exception as e: - # Print errors to stderr so they don't interfere with JSON output - print(json.dumps({"error": str(e)}), file=sys.stderr) - sys.stderr.flush() - + # Log and fall back gracefully + print(json.dumps({"error": str(e)}), file=sys.stderr, flush=True) + return {} class DocumentType(Enum): """ @@ -668,7 +622,7 @@ class Document: Represents a document being processed, such as a PDF, handling chunking, embedding, and summarization. """ - def __init__(self, file_path: str, file_name: str, job_id: str, output_folder: str): + def __init__(self, file_path: str, file_name: str, job_id: str, output_folder: str, doc_id: str): """ Initialize the Document with file data, file name, and job ID. @@ -681,7 +635,7 @@ class Document: self.file_path = file_path self.job_id = job_id self.type = self._get_document_type(file_name) # Determine the document type (PDF, CSV, etc.) - self.doc_id = job_id # Use the job ID as the document ID + self.doc_id = doc_id # Use the job ID as the document ID self.chunks = [] # List to hold text and visual chunks self.num_pages = 0 # Number of pages in the document (if applicable) self.summary = "" # The generated summary for the document @@ -767,7 +721,7 @@ class Document: client = OpenAI() # Initialize OpenAI client for text generation completion = client.chat.completions.create( - model="gpt-3.5-turbo", # Specify the language model + model="gpt-4o", # Specify the language model messages=[ {"role": "system", "content": "You are an AI assistant tasked with summarizing a document. You are provided with important chunks from the document and provide a summary, as best you can, of what the document will contain overall. Be concise and brief with your response."}, @@ -801,7 +755,7 @@ class Document: "doc_id": self.doc_id }, indent=2) # Convert the document's attributes to JSON format -def process_document(file_path, job_id, output_folder): +def process_document(file_path, job_id, output_folder, doc_id): """ Top-level function to process a document and return the JSON output. @@ -809,26 +763,27 @@ def process_document(file_path, job_id, output_folder): :param job_id: The job ID for this document processing task. :return: The processed document's data in JSON format. """ - new_document = Document(file_path, file_path, job_id, output_folder) + new_document = Document(file_path, file_path, job_id, output_folder, doc_id) return new_document.to_json() def main(): """ Main entry point for the script, called with arguments from Node.js. """ - if len(sys.argv) != 4: + if len(sys.argv) != 5: print(json.dumps({"error": "Invalid arguments"}), file=sys.stderr) return job_id = sys.argv[1] file_path = sys.argv[2] output_folder = sys.argv[3] # Get the output folder from arguments + doc_id = sys.argv[4] try: os.makedirs(output_folder, exist_ok=True) # Process the document - document_result = process_document(file_path, job_id, output_folder) # Pass output_folder + document_result = process_document(file_path, job_id, output_folder,doc_id) # Pass output_folder # Output the final result as JSON to stdout print(document_result) |