davebb/linear-coding-agent
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Fix: Métadonnées Word correctes + désactivation concepts

Browse Source 
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>
This commit is contained in:
David Blanc Brioir
2025-12-30 23:39:41 +01:00
parent b928352e36
commit 845ffb4b06
1 changed files with 28 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
generations/library_rag/utils/word_pipeline.py
View File

@@ -239,71 +239,27 @@ def process_word(
) )
# ================================================================ # ================================================================
# STEP 5: LLM Metadata Extraction (REUSED) # 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 metadata: Metadata
cost_llm = 0.0 cost_llm = 0.0
if use_llm:
from utils.llm_metadata import extract_metadata
callback("Metadata Extraction", "running", "Extracting metadata with LLM...")
try:
metadata_llm = extract_metadata(
markdown_text,
provider=llm_provider,
)
# Fallback to Word properties if LLM returns None
if metadata_llm is None:
callback(
"Metadata Extraction",
"completed",
"LLM extraction failed, using Word properties",
)
raw_meta = content["metadata_raw"] raw_meta = content["metadata_raw"]
metadata = Metadata( metadata = Metadata(
title=raw_meta.get("title", doc_name), title=raw_meta.get("title") or doc_name,
author=raw_meta.get("author", "Unknown"), author=raw_meta.get("author") or "Unknown",
year=raw_meta.get("created").year if raw_meta.get("created") else None, year=raw_meta.get("created").year if raw_meta.get("created") else None,
language=raw_meta.get("language", "unknown"), language="en", # Default to English, could be improved
) )
else:
metadata = metadata_llm
callback( callback(
"Metadata Extraction", "Metadata Extraction",
"completed", "completed",
f"Title: {metadata.get('title', '')[:50]}..., Author: {metadata.get('author', '')}", f"Title: {metadata.get('title', '')[:50]}..., Author: {metadata.get('author', '')}",
) )
except Exception as e:
callback(
"Metadata Extraction",
"completed",
f"LLM error ({str(e)}), using Word properties",
)
raw_meta = content["metadata_raw"]
metadata = Metadata(
title=raw_meta.get("title", doc_name),
author=raw_meta.get("author", "Unknown"),
year=raw_meta.get("created").year if raw_meta.get("created") else None,
language=raw_meta.get("language", "unknown"),
)
else:
# Use metadata from Word properties
raw_meta = content["metadata_raw"]
metadata = Metadata(
title=raw_meta.get("title", doc_name),
author=raw_meta.get("author", "Unknown"),
year=raw_meta.get("created").year if raw_meta.get("created") else None,
language=raw_meta.get("language", "unknown"),
)
callback(
"Metadata Extraction",
"completed",
"Using Word document properties",
)
# ================================================================ # ================================================================
# STEP 6: Section Classification (REUSED) # STEP 6: Section Classification (REUSED)
@@ -452,25 +408,18 @@ def process_word(
) )
# ================================================================ # ================================================================
# STEP 9: Chunk Validation (REUSED) # STEP 9: Chunk Validation (SKIP FOR WORD)
# ================================================================ # ================================================================
if use_llm: # NOTE: We skip LLM concept enrichment for Word documents because:
from utils.llm_validator import enrich_chunks_with_concepts # 1. The LLM generates concepts in French even for English text
# 2. French accents cause Weaviate ingestion failures
callback("Chunk Validation", "running", "Enriching chunks with concepts...") # 3. Word documents already have clean structure, don't need LLM enhancement
#
# Enrich chunks with keywords/concepts # For production: could re-enable with language detection + prompt tuning
enriched_chunks = enrich_chunks_with_concepts(
chunks,
provider=llm_provider,
)
chunks = enriched_chunks
callback( callback(
"Chunk Validation", "Chunk Validation",
"completed", "completed",
f"Validated {len(chunks)} chunks", f"Skipped (Word documents don't need LLM enrichment)",
) )
# ================================================================ # ================================================================