davebb/linear-coding-agent
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity
Files
845ffb4b06f3840a2e14ecd610753772d5c373ab
linear-coding-agent/generations/library_rag/utils/word_pipeline.py
David Blanc Brioir 845ffb4b06 Fix: Métadonnées Word correctes + désactivation concepts 
Problèmes corrigés:
1. TITRE INCORRECT → Maintenant utilise TITRE: de la première page
2. CONCEPTS EN FRANÇAIS → Désactivé l'enrichissement LLM

Avant:
- Titre: "An Historical Sketch..." (mauvais, titre du chapitre)
- Concepts: ['immuabilité des espèces', 'création séparée'] (français)
- Résultat: 3/37 chunks ingérés dans Weaviate

Après:
- Titre: "On the Origin of Species BY MEANS OF..." (correct!)
- Concepts: [] (vides, pas de problème d'encoding)
- Résultat: 14/37 chunks ingérés (mieux mais pas parfait)

Changements word_pipeline.py:

1. STEP 5 - Métadonnées simplifiées (ligne 241-262):
   - Supprimé l'appel à extract_metadata() du LLM
   - Utilise directement raw_meta de extract_word_metadata()
   - Le LLM prenait le titre du chapitre au lieu du livre

2. STEP 9 - Désactivé enrichissement concepts (ligne 410-423):
   - Skip enrich_chunks_with_concepts()
   - Raison: LLM génère concepts en FRANÇAIS pour texte ANGLAIS
   - Accents français causent échecs Weaviate

Note TOC:
Le document n'a que 2 Heading 2, donc la TOC est limitée.
C'est normal pour un extrait de 10 pages.

Reste à investiguer: Pourquoi 14/37 au lieu de 37/37 chunks?

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
2025-12-30 23:39:41 +01:00

509 lines
18 KiB
Python
Raw Blame History

"""Word document processing pipeline for RAG ingestion.
This module provides a complete pipeline for processing Microsoft Word documents
(.docx) through the RAG system. It extracts content, builds structured markdown,
applies LLM processing, and ingests chunks into Weaviate.
The pipeline reuses existing LLM modules (metadata extraction, classification,
chunking, cleaning, validation) from the PDF pipeline, only replacing the initial
extraction step with Word-specific processing.
Example:
Process a Word document with default settings:
from pathlib import Path
from utils.word_pipeline import process_word
result = process_word(
Path("document.docx"),
use_llm=True,
llm_provider="ollama",
ingest_to_weaviate=True,
)
print(f"Success: {result['success']}")
print(f"Chunks created: {result['chunks_count']}")
Process without Weaviate ingestion:
result = process_word(
Path("document.docx"),
use_llm=True,
ingest_to_weaviate=False,
)
Pipeline Steps:
1. Word Extraction (word_processor.py)
2. Markdown Construction
3. TOC Extraction (word_toc_extractor.py)
4. Metadata Extraction (llm_metadata.py) - REUSED
5. Section Classification (llm_classifier.py) - REUSED
6. Semantic Chunking (llm_chunker.py) - REUSED
7. Chunk Cleaning (llm_cleaner.py) - REUSED
8. Chunk Validation (llm_validator.py) - REUSED
9. Weaviate Ingestion (weaviate_ingest.py) - REUSED
See Also:
- utils.word_processor: Word content extraction
- utils.word_toc_extractor: TOC construction from headings
- utils.pdf_pipeline: Similar pipeline for PDF documents
"""
from pathlib import Path
from typing import Any, Dict, List, Optional, Callable
import json
from utils.types import (
Metadata,
TOCEntry,
ChunkData,
PipelineResult,
LLMProvider,
ProgressCallback,
)
from utils.word_processor import (
extract_word_content,
extract_word_metadata,
build_markdown_from_word,
extract_word_images,
)
from utils.word_toc_extractor import build_toc_from_headings, flatten_toc
# Note: LLM modules imported dynamically when use_llm=True to avoid import errors
def _default_progress_callback(step: str, status: str, detail: str = "") -> None:
"""Default progress callback that prints to console.
Args:
step: Current pipeline step name.
status: Step status (running, completed, error).
detail: Optional detail message.
"""
status_symbol = {
"running": ">>>",
"completed": "[OK]",
"error": "[ERROR]",
}.get(status, "[INFO]")
print(f"{status_symbol} {step}: {detail}" if detail else f"{status_symbol} {step}")
def process_word(
word_path: Path,
*,
use_llm: bool = True,
llm_provider: LLMProvider = "ollama",
use_semantic_chunking: bool = True,
ingest_to_weaviate: bool = True,
skip_metadata_lines: int = 5,
extract_images: bool = True,
progress_callback: Optional[ProgressCallback] = None,
) -> PipelineResult:
"""Process a Word document through the complete RAG pipeline.
Extracts content from a .docx file, processes it with LLM modules,
and optionally ingests the chunks into Weaviate. Reuses all LLM
processing steps from the PDF pipeline (metadata, classification,
chunking, cleaning, validation).
Args:
word_path: Path to the .docx file to process.
use_llm: Enable LLM processing steps (metadata, chunking, validation).
If False, uses simple text splitting. Default: True.
llm_provider: LLM provider to use ("ollama" for local, "mistral" for API).
Default: "ollama".
use_semantic_chunking: Use LLM-based semantic chunking instead of simple
text splitting. Requires use_llm=True. Default: True.
ingest_to_weaviate: Ingest processed chunks into Weaviate database.
Default: True.
skip_metadata_lines: Number of initial paragraphs to skip when building
markdown (metadata header lines like TITRE, AUTEUR). Default: 5.
extract_images: Extract and save inline images from the document.
Default: True.
progress_callback: Optional callback for progress updates.
Signature: (step: str, status: str, detail: str) -> None.
Returns:
PipelineResult dictionary with keys:
- success (bool): Whether processing succeeded
- document_name (str): Name of processed document
- output_dir (Path): Directory containing outputs
- chunks_count (int): Number of chunks created
- cost_ocr (float): OCR cost (always 0 for Word)
- cost_llm (float): LLM processing cost
- cost_total (float): Total cost
- error (str): Error message if success=False
Raises:
FileNotFoundError: If word_path does not exist.
ValueError: If file is not a .docx document.
Example:
>>> result = process_word(
... Path("darwin.docx"),
... use_llm=True,
... llm_provider="ollama",
... ingest_to_weaviate=True,
... )
>>> print(f"Created {result['chunks_count']} chunks")
>>> print(f"Total cost: ${result['cost_total']:.4f}")
Note:
No OCR cost for Word documents (cost_ocr always 0).
LLM costs depend on provider and document length.
"""
# Use default progress callback if none provided
callback = progress_callback or _default_progress_callback
try:
# Validate input
if not word_path.exists():
raise FileNotFoundError(f"Word document not found: {word_path}")
if not word_path.suffix.lower() == ".docx":
raise ValueError(f"File must be .docx format: {word_path}")
doc_name = word_path.stem
output_dir = Path("output") / doc_name
output_dir.mkdir(parents=True, exist_ok=True)
# ================================================================
# STEP 1: Extract Word Content
# ================================================================
callback("Word Extraction", "running", "Extracting document content...")
content = extract_word_content(word_path)
callback(
"Word Extraction",
"completed",
f"Extracted {content['total_paragraphs']} paragraphs, "
f"{len(content['headings'])} headings",
)
# ================================================================
# STEP 2: Build Markdown
# ================================================================
callback("Markdown Construction", "running", "Building markdown...")
markdown_text = build_markdown_from_word(
content["paragraphs"],
skip_metadata_lines=skip_metadata_lines,
)
# Save markdown
markdown_path = output_dir / f"{doc_name}.md"
with open(markdown_path, "w", encoding="utf-8") as f:
f.write(markdown_text)
callback(
"Markdown Construction",
"completed",
f"Saved to {markdown_path.name} ({len(markdown_text)} chars)",
)
# ================================================================
# STEP 3: Build TOC
# ================================================================
callback("TOC Extraction", "running", "Building table of contents...")
toc_hierarchical = build_toc_from_headings(content["headings"])
toc_flat = flatten_toc(toc_hierarchical)
callback(
"TOC Extraction",
"completed",
f"Built TOC with {len(toc_flat)} entries",
)
# ================================================================
# STEP 4: Extract Images (if requested)
# ================================================================
image_paths: List[Path] = []
if extract_images and content["has_images"]:
callback("Image Extraction", "running", "Extracting images...")
from docx import Document
doc = Document(word_path)
image_paths = extract_word_images(
doc,
output_dir / "images",
doc_name,
)
callback(
"Image Extraction",
"completed",
f"Extracted {len(image_paths)} images",
)
# ================================================================
# STEP 5: Metadata Extraction from Word (NO LLM NEEDED)
# ================================================================
# Word documents have metadata in first lines (TITRE:, AUTEUR:, EDITION:)
# or in core properties. LLM extraction often gets it wrong (takes chapter
# title instead of book title), so we use Word-native metadata directly.
metadata: Metadata
cost_llm = 0.0
raw_meta = content["metadata_raw"]
metadata = Metadata(
title=raw_meta.get("title") or doc_name,
author=raw_meta.get("author") or "Unknown",
year=raw_meta.get("created").year if raw_meta.get("created") else None,
language="en", # Default to English, could be improved
)
callback(
"Metadata Extraction",
"completed",
f"Title: {metadata.get('title', '')[:50]}..., Author: {metadata.get('author', '')}",
)
# ================================================================
# STEP 6: Section Classification (REUSED)
# ================================================================
if use_llm:
from utils.llm_classifier import classify_sections
callback("Section Classification", "running", "Classifying sections...")
# Note: classify_sections expects a list of section dicts, not raw TOC
sections_to_classify = [
{
"section_path": entry["sectionPath"],
"title": entry["title"],
"content": "", # Content matched later
}
for entry in toc_flat
]
classified_sections = classify_sections(
sections_to_classify,
document_title=metadata.get("title", ""),
provider=llm_provider,
)
main_sections = [
s for s in classified_sections
if s.get("type") == "main_content"
]
callback(
"Section Classification",
"completed",
f"{len(main_sections)}/{len(classified_sections)} main content sections",
)
else:
# All sections are main content by default
classified_sections = [
{
"section_path": entry["sectionPath"],
"type": "main_content",
"should_index": True,
"classification_reason": "No LLM classification",
}
for entry in toc_flat
]
callback(
"Section Classification",
"completed",
"Skipped (use_llm=False)",
)
# ================================================================
# STEP 7: Semantic Chunking (REUSED)
# ================================================================
if use_llm and use_semantic_chunking:
from utils.llm_chunker import chunk_section_with_llm
callback("Semantic Chunking", "running", "Chunking with LLM...")
# Chunk each section
all_chunks: List[ChunkData] = []
for entry in toc_flat:
# TODO: Extract section content from markdown based on sectionPath
# For now, using simple approach
section_chunks = chunk_section_with_llm(
markdown_text,
entry["title"],
metadata.get("title", ""),
metadata.get("author", ""),
provider=llm_provider,
)
all_chunks.extend(section_chunks)
chunks = all_chunks
callback(
"Semantic Chunking",
"completed",
f"Created {len(chunks)} semantic chunks",
)
else:
# Simple text splitting (fallback)
callback("Text Splitting", "running", "Simple text splitting...")
# Simple chunking by paragraphs (basic fallback)
chunks_simple = []
for i, para in enumerate(content["paragraphs"][skip_metadata_lines:]):
if para["text"] and not para["is_heading"]:
chunk_dict: ChunkData = {
"text": para["text"],
"keywords": [],
"sectionPath": "1", # Default section
"chapterTitle": "Main Content",
"unitType": "paragraph",
"orderIndex": i,
"work": {
"title": metadata["title"],
"author": metadata["author"],
},
"document": {
"sourceId": doc_name,
"edition": content["metadata_raw"].get("edition", ""),
},
}
chunks_simple.append(chunk_dict)
chunks = chunks_simple
callback(
"Text Splitting",
"completed",
f"Created {len(chunks)} simple chunks",
)
# ================================================================
# STEP 8: Chunk Cleaning (REUSED)
# ================================================================
if use_llm:
from utils.llm_cleaner import clean_chunk, is_chunk_valid
callback("Chunk Cleaning", "running", "Cleaning chunks...")
# Clean each chunk
cleaned_chunks = []
for chunk in chunks:
# Extract text from chunk dict
text: str = chunk.get("text", "")
# Clean the text
cleaned_text = clean_chunk(text, use_llm=False)
# Validate chunk
if is_chunk_valid(cleaned_text, min_chars=30, min_words=8):
# Update chunk with cleaned text
chunk["text"] = cleaned_text
cleaned_chunks.append(chunk)
chunks = cleaned_chunks
callback(
"Chunk Cleaning",
"completed",
f"{len(chunks)} chunks after cleaning",
)
# ================================================================
# STEP 9: Chunk Validation (SKIP FOR WORD)
# ================================================================
# NOTE: We skip LLM concept enrichment for Word documents because:
# 1. The LLM generates concepts in French even for English text
# 2. French accents cause Weaviate ingestion failures
# 3. Word documents already have clean structure, don't need LLM enhancement
#
# For production: could re-enable with language detection + prompt tuning
callback(
"Chunk Validation",
"completed",
f"Skipped (Word documents don't need LLM enrichment)",
)
# ================================================================
# STEP 10: Save Chunks JSON
# ================================================================
callback("Save Results", "running", "Saving chunks to JSON...")
chunks_output = {
"metadata": metadata,
"toc": toc_flat,
"classified_sections": classified_sections,
"chunks": chunks,
"cost_ocr": 0.0, # No OCR for Word documents
"cost_llm": cost_llm,
"cost_total": cost_llm,
"paragraphs": content["total_paragraphs"],
"chunks_count": len(chunks),
}
chunks_path = output_dir / f"{doc_name}_chunks.json"
with open(chunks_path, "w", encoding="utf-8") as f:
json.dump(chunks_output, f, indent=2, ensure_ascii=False, default=str)
callback(
"Save Results",
"completed",
f"Saved to {chunks_path.name}",
)
# ================================================================
# STEP 11: Weaviate Ingestion (REUSED)
# ================================================================
if ingest_to_weaviate:
from utils.weaviate_ingest import ingest_document
callback("Weaviate Ingestion", "running", "Ingesting into Weaviate...")
ingestion_result = ingest_document(
doc_name=doc_name,
chunks=chunks,
metadata=metadata,
language=metadata.get("language", "unknown"),
toc=toc_flat,
hierarchy=None, # Word documents don't have page-based hierarchy
pages=0, # Word documents don't have pages
)
# Save ingestion results
weaviate_path = output_dir / f"{doc_name}_weaviate.json"
with open(weaviate_path, "w", encoding="utf-8") as f:
json.dump(ingestion_result, f, indent=2, ensure_ascii=False, default=str)
callback(
"Weaviate Ingestion",
"completed",
f"Ingested {ingestion_result.get('count', 0)} chunks",
)
# ================================================================
# Return Success Result
# ================================================================
return PipelineResult(
success=True,
document_name=doc_name,
output_dir=output_dir,
chunks_count=len(chunks),
cost_ocr=0.0,
cost_llm=cost_llm,
cost_total=cost_llm,
error="",
)
except Exception as e:
error_msg = f"Pipeline failed: {str(e)}"
callback("Pipeline Error", "error", error_msg)
return PipelineResult(
success=False,
document_name=word_path.stem,
output_dir=Path("output") / word_path.stem,
chunks_count=0,
cost_ocr=0.0,
cost_llm=0.0,
cost_total=0.0,
error=error_msg,
)
Reference in New Issue View Git Blame Copy Permalink