.env.example

# .env.example — extractor pipeline

# LiteLLM models
LITELLM_DEFAULT_MODEL=openai/gpt-5
LITELLM_VLM_MODEL=openai/gpt-4o
MAX_CONCURRENT_LLM_CALLS=3

# Stage 05 — Table extraction (Camelot/pandas)
TABLE_VERTICAL_PADDING_RATIO=0.30
TABLE_HORIZONTAL_PADDING_RATIO=0.07
TABLE_EXTRACTION_DPI=200
TABLE_STITCH_MIN_HORIZONTAL_IOU=0.2
TABLE_STITCH_ALLOW_NEXT_PAGE=true
TABLE_FILTER_MIN_DENSITY=0.15
TABLE_FILTER_MIN_ROWS=3
TABLE_HEADER_DUP_MIN_MATCH=0.5
TABLE_HEADER_COALESCE_ENABLED=true
TABLE_HEADER_REPEAT_MIN_MATCH=0.6
TABLE_MULTI_PAGE_MERGE_ENABLED=true
TABLE_MULTI_PAGE_MERGE_MIN_IOU=0.3

# Stage 06 — Figure extraction
FIGURE_VERTICAL_PADDING=0.10

# Stage 07 — Reflow + hybrid search
SEMANTIC_ANNOTATION_TOP_K=5
ARANGO_ENABLE_HYBRID=true
ARANGO_HOST=localhost
ARANGO_PORT=8529
ARANGO_USER=root
ARANGO_PASSWORD=changeme
ARANGO_DATABASE=pdf_knowledge_base
ARANGO_ANNOTATIONS_COLLECTION=
ARANGO_EDGES_COLLECTION=

# Stage 10/11 — Embeddings/graph
EMBEDDING_MODEL=sentence-transformers/all-MiniLM-L6-v2

# Notes:
# - Copy to .env and customize.
# - For security, do not commit your .env.
# - Install camelot-py[cv] and ghostscript for Stage 05.

# Stage 11 — Graph/FAISS relationship building
GRAPH_K_NEIGHBORS=10
GRAPH_SIMILARITY_THRESHOLD=0.55
GRAPH_SEMANTIC_WEIGHT=0.7
GRAPH_HIERARCHY_WEIGHT=0.3
GRAPH_EDGE_COLLECTION=pdf_relationships
GRAPH_NAME=pdf_knowledge_graph
GRAPH_VERTEX_COLLECTION=pdf_objects
GRAPH_ENABLE_RATIONALES=true
GRAPH_RATIONALE_MODEL=openai/gpt-5-mini
GRAPH_RATIONALE_CONCURRENCY=8
GRAPH_RATIONALE_MAX_TOKENS=256