| import json | |
| import logging | |
| import uuid | |
| import math | |
| import os | |
| import sys | |
| from dataclasses import dataclass, field | |
| from typing import List, Dict, Any, Set, Optional | |
| # ----------------------------------------------------------------------------- | |
| # Imports & Dependency Checks | |
| # ----------------------------------------------------------------------------- | |
| try: | |
| import yaml | |
| except ImportError: | |
| print("Error: 'PyYAML' is required. Install via 'pip install pyyaml'.") | |
| sys.exit(1) | |
| try: | |
| from openai import OpenAI, OpenAIError | |
| except ImportError: | |
| print("Error: 'openai' is required. Install via 'pip install openai'.") | |
| sys.exit(1) | |
| # We check for transformers inside the class to avoid crashing if | |
| # the user wants heuristic mode but doesn't have transformers installed. | |
| try: | |
| from transformers import AutoTokenizer | |
| TRANSFORMERS_AVAILABLE = True | |
| except ImportError: | |
| TRANSFORMERS_AVAILABLE = False | |
| # ----------------------------------------------------------------------------- | |
| # Logging | |
| # ----------------------------------------------------------------------------- | |
| logging.basicConfig( | |
| level=logging.DEBUG, | |
| format='[%(levelname)s] %(asctime)s - %(funcName)s:%(lineno)d - %(message)s', | |
| datefmt='%H:%M:%S' | |
| ) | |
| logger = logging.getLogger(__name__) | |
| # ----------------------------------------------------------------------------- | |
| # Configuration | |
| # ----------------------------------------------------------------------------- | |
| @dataclass | |
| class GroupInterval: | |
| start: int | |
| end: int | |
| line_numbers: Set[int] | |
| @dataclass | |
| class ChunkingConfig: | |
| """Configuration object loaded from YAML.""" | |
| api_key: str | |
| llm_model_name: str | |
| temperature: float | |
| # Tokenization | |
| tokenizer_method: str | |
| hf_model_name: str | |
| heuristic_chars_per_token: int | |
| # Limits | |
| llm_token_limit: int | |
| overlap_token_count: int | |
| model_token_limit: int | |
| # Prompts | |
| system_prompt_base: str | |
| @classmethod | |
| def from_yaml(cls, path: str) -> 'ChunkingConfig': | |
| if not os.path.exists(path): | |
| raise FileNotFoundError(f"Config file not found at: {path}") | |
| logger.info(f"Loading configuration from {path}...") | |
| with open(path, 'r') as f: | |
| data = yaml.safe_load(f) | |
| oa = data.get('openai', {}) | |
| tok = data.get('tokenization', {}) | |
| tok_heu = tok.get('heuristic', {}) | |
| tok_hf = tok.get('huggingface', {}) | |
| lim = data.get('limits', {}) | |
| prompts = data.get('prompts', {}) | |
| raw_key = oa.get('api_key', 'ENV') | |
| api_key = os.getenv("OPENAI_API_KEY") if raw_key == "ENV" else raw_key | |
| return cls( | |
| api_key=api_key or "MISSING_KEY", | |
| llm_model_name=oa.get('model_name', 'gpt-4o-mini'), | |
| temperature=oa.get('temperature', 0.0), | |
| # Tokenizer Config | |
| tokenizer_method=tok.get('method', 'heuristic'), | |
| hf_model_name=tok_hf.get('model_name', 'gpt2'), | |
| heuristic_chars_per_token=tok_heu.get('chars_per_token', 4), | |
| llm_token_limit=lim.get('llm_context_window', 300), | |
| overlap_token_count=lim.get('window_overlap', 50), | |
| model_token_limit=lim.get('target_chunk_size', 100), | |
| system_prompt_base=prompts.get('system_instructions', '') | |
| ) | |
| # ----------------------------------------------------------------------------- | |
| # Data Structures | |
| # ----------------------------------------------------------------------------- | |
| @dataclass | |
| class Line: | |
| number: int | |
| text: str | |
| token_count: int | |
| @dataclass | |
| class PreChunkSegment: | |
| lines: List[Line] | |
| segment_id: str = field(default_factory=lambda: str(uuid.uuid4())) | |
| @property | |
| def formatted_text(self) -> str: | |
| return "\n".join([f"{line.number} | {line.text}" for line in self.lines]) | |
| @dataclass | |
| class SemanticGroup: | |
| line_numbers: Set[int] | |
| # ----------------------------------------------------------------------------- | |
| # Service Implementation | |
| # ----------------------------------------------------------------------------- | |
| class DocumentChunkingService: | |
| def __init__(self, config_path: str = "config.yaml"): | |
| # 1. Load Config | |
| try: | |
| self.config = ChunkingConfig.from_yaml(config_path) | |
| except Exception as e: | |
| logger.critical(f"Failed to load config: {e}") | |
| sys.exit(1) | |
| # 2. Setup Tokenizer based on Method | |
| self.hf_tokenizer = None | |
| if self.config.tokenizer_method == "huggingface": | |
| if not TRANSFORMERS_AVAILABLE: | |
| logger.critical("Config requests 'huggingface', but library is missing. Install 'transformers'.") | |
| sys.exit(1) | |
| try: | |
| logger.info(f"Initializing HuggingFace Tokenizer: {self.config.hf_model_name}") | |
| os.environ["TOKENIZERS_PARALLELISM"] = "false" | |
| self.hf_tokenizer = AutoTokenizer.from_pretrained(self.config.hf_model_name) | |
| except Exception as e: | |
| logger.critical(f"Failed to load HF Tokenizer: {e}") | |
| sys.exit(1) | |
| elif self.config.tokenizer_method == "heuristic": | |
| logger.info(f"Using Heuristic Tokenizer ({self.config.heuristic_chars_per_token} chars/token)") | |
| else: | |
| logger.warning(f"Unknown tokenizer method '{self.config.tokenizer_method}'. Defaulting to heuristic.") | |
| # 3. Setup OpenAI | |
| if self.config.api_key == "MISSING_KEY": | |
| logger.critical("No valid API Key found.") | |
| self.client = None | |
| else: | |
| try: | |
| self.client = OpenAI(api_key=self.config.api_key) | |
| except Exception as e: | |
| logger.error(f"Failed to initialize OpenAI Client: {e}") | |
| sys.exit(1) | |
| def _count_tokens(self, text: str) -> int: | |
| """ | |
| Determines token count based on the configured method. | |
| """ | |
| if not text: | |
| return 0 | |
| if self.config.tokenizer_method == "huggingface" and self.hf_tokenizer: | |
| # HuggingFace Count | |
| return len(self.hf_tokenizer.encode(text, add_special_tokens=False)) | |
| else: | |
| # Heuristic Count | |
| return math.ceil(len(text) / self.config.heuristic_chars_per_token) | |
| def _prepare_lines(self, document_text: str) -> List[Line]: | |
| logger.debug(f"Preparing lines using {self.config.tokenizer_method} method...") | |
| raw_lines = document_text.split('\n') | |
| processed_lines = [] | |
| for idx, text in enumerate(raw_lines, start=1): | |
| if not text.strip(): continue | |
| count = self._count_tokens(text) | |
| processed_lines.append(Line(idx, text, count)) | |
| return processed_lines | |
| def _create_pre_chunks(self, lines: List[Line]) -> List[PreChunkSegment]: | |
| logger.debug(f"Segmenting lines (Limit: {self.config.llm_token_limit})...") | |
| segments = [] | |
| current_segment_lines = [] | |
| current_tokens = 0 | |
| i = 0 | |
| while i < len(lines): | |
| line = lines[i] | |
| if current_tokens + line.token_count > self.config.llm_token_limit and current_segment_lines: | |
| segments.append(PreChunkSegment(list(current_segment_lines))) | |
| # Overlap Logic | |
| overlap_buffer = [] | |
| overlap_tokens = 0 | |
| back_idx = i - 1 | |
| while back_idx >= 0: | |
| prev_line = lines[back_idx] | |
| if prev_line in current_segment_lines: | |
| overlap_buffer.insert(0, prev_line) | |
| overlap_tokens += prev_line.token_count | |
| if overlap_tokens >= self.config.overlap_token_count: | |
| break | |
| else: | |
| break | |
| back_idx -= 1 | |
| current_segment_lines = list(overlap_buffer) | |
| current_tokens = overlap_tokens | |
| current_segment_lines.append(line) | |
| current_tokens += line.token_count | |
| i += 1 | |
| if current_segment_lines: | |
| segments.append(PreChunkSegment(current_segment_lines)) | |
| return segments | |
| def _call_openai(self, segment_text: str, available_lines: List[int]) -> List[List[int]]: | |
| runtime_constraint = f"\nCRITICAL CONSTRAINT: Only use the line numbers provided in this specific range: {available_lines}" | |
| full_system_prompt = self.config.system_prompt_base + runtime_constraint | |
| user_prompt = f"Input Lines:\n{segment_text}\n\nOutput JSON:" | |
| try: | |
| logger.debug(f"Calling OpenAI (Lines {available_lines[0]}-{available_lines[-1]})...") | |
| response = self.client.chat.completions.create( | |
| model=self.config.llm_model_name, | |
| messages=[ | |
| {"role": "system", "content": full_system_prompt}, | |
| {"role": "user", "content": user_prompt} | |
| ], | |
| response_format={"type": "json_object"}, | |
| temperature=self.config.temperature | |
| ) | |
| parsed = json.loads(response.choices[0].message.content) | |
| groups = parsed.get("groups", []) | |
| if isinstance(groups, list) and all(isinstance(g, list) for g in groups): | |
| return groups | |
| return [[l] for l in available_lines] | |
| except Exception as e: | |
| logger.error(f"OpenAI Call Failed: {e}") | |
| return [[l] for l in available_lines] | |
| def _get_semantic_groupings(self, segments: List[PreChunkSegment]) -> List[List[int]]: | |
| all_raw_groups = [] | |
| for idx, seg in enumerate(segments): | |
| available_lines = [l.number for l in seg.lines] | |
| response_groups = self._call_openai(seg.formatted_text, available_lines) | |
| all_raw_groups.extend(response_groups) | |
| return all_raw_groups | |
| def resolve_overlaps(raw_groups: List[List[int]], all_lines_map: Dict[int, Line]) -> List[SemanticGroup]: | |
| """ | |
| Merges groups based on overlapping line number ranges. | |
| Uses a standard 'Merge Intervals' algorithm. | |
| """ | |
| intervals: List[GroupInterval] = [] | |
| # 1. Convert raw groups to Intervals | |
| for group in raw_groups: | |
| if not group: | |
| continue | |
| # Filter for valid lines only | |
| valid_lines = {g for g in group if g in all_lines_map} | |
| if not valid_lines: | |
| continue | |
| # Define range based on min and max line numbers in the group | |
| intervals.append(GroupInterval( | |
| start=min(valid_lines), | |
| end=max(valid_lines), | |
| line_numbers=valid_lines | |
| )) | |
| if not intervals: | |
| return [] | |
| # 2. Sort by start time | |
| intervals.sort(key=lambda x: x.start) | |
| # 3. Merge overlapping intervals | |
| merged: List[GroupInterval] = [] | |
| for current in intervals: | |
| if not merged: | |
| merged.append(current) | |
| continue | |
| last = merged[-1] | |
| # Check for overlap: | |
| # If current starts before (or exactly when) last ends, they overlap. | |
| # e.g. [84, 795] and [788, 887] -> 788 <= 795, so merge. | |
| if current.start <= last.end: | |
| # Merge logic: | |
| # 1. Extend the end if needed | |
| last.end = max(last.end, current.end) | |
| # 2. Combine the sets of line numbers | |
| last.line_numbers.update(current.line_numbers) | |
| else: | |
| # No overlap, start a new cluster | |
| merged.append(current) | |
| # 4. Convert back to SemanticGroups | |
| results = [SemanticGroup(group.line_numbers) for group in merged] | |
| return sorted(results, key=lambda x: min(x.line_numbers) if x.line_numbers else 0) | |
| def _finalize_chunk(self, content: str, line_numbers: List[int], parent_id: Optional[str] = None) -> List[Dict[str, Any]]: | |
| count = self._count_tokens(content) | |
| if count <= self.config.model_token_limit: | |
| return [{ | |
| "content": content, | |
| "line_numbers": line_numbers, | |
| "token_estimate": count, | |
| "metadata": {"parent_id": parent_id} | |
| }] | |
| if len(line_numbers) <= 1: | |
| return [{ | |
| "content": content, | |
| "line_numbers": line_numbers, | |
| "token_estimate": count, | |
| "metadata": {"parent_id": parent_id, "warning": "oversized"} | |
| }] | |
| mid = len(line_numbers) // 2 | |
| left_lines = line_numbers[:mid] | |
| right_lines = line_numbers[mid:] | |
| left_text = "\n".join([self.current_doc_map[n].text for n in left_lines]) | |
| right_text = "\n".join([self.current_doc_map[n].text for n in right_lines]) | |
| cid = parent_id if parent_id else str(uuid.uuid4())[:8] | |
| results = [] | |
| results.extend(self._finalize_chunk(left_text, left_lines, parent_id=cid)) | |
| results.extend(self._finalize_chunk(right_text, right_lines, parent_id=cid)) | |
| return results | |
| def process_document(self, plaintext: str) -> str: | |
| logger.info(f">>> Processing Document [Mode: {self.config.tokenizer_method.upper()}]") | |
| lines = self._prepare_lines(plaintext) | |
| self.current_doc_map = {l.number: l for l in lines} | |
| pre_chunks = self._create_pre_chunks(lines) | |
| raw_groups = self._get_semantic_groupings(pre_chunks) | |
| merged_groups = self._resolve_overlaps(raw_groups, self.current_doc_map) | |
| final_output = [] | |
| logger.info("Finalizing chunks...") | |
| for group in merged_groups: | |
| sorted_nums = sorted(list(group.line_numbers)) | |
| text_content = "\n".join([self.current_doc_map[n].text for n in sorted_nums]) | |
| chunks = self._finalize_chunk(text_content, sorted_nums) | |
| final_output.extend(chunks) | |
| logger.info(f"<<< Done. Generated {len(final_output)} chunks.") | |
| return json.dumps(final_output, indent=2) | |
| # ----------------------------------------------------------------------------- | |
| # Main Execution | |
| # ----------------------------------------------------------------------------- | |
| if __name__ == "__main__": | |
| sample_text = """The history of Artificial Intelligence is fascinating. | |
| It begins with the Turing Test proposed by Alan Turing. | |
| Early AI research focused on symbolic logic and problem solving. | |
| However, computing power was limited in the 1950s. | |
| Decades later, machine learning emerged as a dominant paradigm. | |
| Neural networks, inspired by the human brain, gained popularity. | |
| Deep learning revolutionized the field in the 2010s. | |
| Transformers, introduced by Google, changed NLP forever. | |
| Large Language Models like GPT-4 are now commonplace. | |
| Retrieval Augmented Generation allows LLMs to use external data. | |
| Chunking documents is essential for RAG systems. | |
| It preserves semantic meaning during retrieval. | |
| This specific code implements a rigorous chunking strategy. | |
| It uses heuristic strategies for token estimation. | |
| The end goal is high quality embeddings.""" | |
| service = DocumentChunkingService("config.yaml") | |
| if service.client: | |
| result = service.process_document(sample_text) | |
| print("\n--- Final Output JSON ---") | |
| print(result) | |
| openai: | |
| api_key: "ENV" | |
| model_name: "gpt-4o-mini" | |
| temperature: 0.0 | |
| tokenization: | |
| # MASTER SWITCH: Choose "heuristic" or "huggingface" | |
| # - "heuristic": Uses simple math (chars / chars_per_token). Fast, no dependencies. | |
| # - "huggingface": Uses a real tokenizer (e.g., gpt2). Precise, requires 'transformers' lib. | |
| method: "heuristic" | |
| # Settings for "heuristic" method | |
| heuristic: | |
| chars_per_token: 4 | |
| # Settings for "huggingface" method | |
| huggingface: | |
| # "gpt2" is a standard proxy for general LLM token counting | |
| model_name: "gpt2" | |
| limits: | |
| # Max tokens to send to OpenAI in one request (chunk context window) | |
| llm_context_window: 300 | |
| # Overlap between context windows to prevent cutting sentences | |
| window_overlap: 50 | |
| # The target max size for a final, atomic chunk | |
| target_chunk_size: 100 | |
| prompts: | |
| system_instructions: | | |
| You are a document chunking assistant. Your goal is to group lines of text into semantically coherent chunks. | |
| Strict Rules: | |
| 1. Every line number provided in the input must appear exactly once in your output. | |
| 2. Group line numbers that belong together conceptually. | |
| 3. Return a JSON object with a single key 'groups' containing a list of lists of integers. | |