.env.example

# Copy this file to `.env` (repo root).
# Priority: shell `export` > `.env` > `i2p_config.json` > code defaults.
#
# IMPORTANT:
# - Do NOT commit `.env` (it is gitignored).
# - Boolean values in env are treated as: only `"1"` is true, everything else is false.
#   (So use 1/0, not true/false.)
#
# -----------------------------
# LLM API Key (secret)
# -----------------------------
# Used for LLM calls only. Do not put in i2p_config.json.
LLM_API_KEY=your_key_here

# -----------------------------
# LLM (provider selection)
# -----------------------------
# Provider options:
# - openai_compatible_chat
# - openai_responses
# - anthropic
# - gemini
LLM_PROVIDER=openai_compatible_chat

# Optional: base URL (recommended)
# For OpenAI-compatible Chat:
#   `https://api.openai.com/v1` (default)
# For other compatible services:
#   `https://dashscope.aliyuncs.com/compatible-mode/v1`
LLM_BASE_URL=https://api.openai.com/v1

# Optional: full API URL (highest priority override)
# Example (OpenAI-compatible Chat):
#   `https://api.openai.com/v1/chat/completions`
# Example (OpenAI Responses):
#   `https://api.openai.com/v1/responses`
# Example (Anthropic):
#   `https://api.anthropic.com/v1/messages`
# Example (Gemini):
#   `https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent`
LLM_API_URL=https://api.openai.com/v1/chat/completions

# Model (provider-specific)
LLM_MODEL=claude-opus-4-5-20251101

# Anthropic optional version override
LLM_ANTHROPIC_VERSION=2025-11-01

# Optional: extra headers/body as JSON (advanced use)
# LLM_EXTRA_HEADERS_JSON={"x-foo":"bar"}
# LLM_EXTRA_BODY_JSON={"top_p":0.9}

# Optional: per-stage LLM temperatures (defaults preserve current behavior)
# Critic is usually low temp for stability; story generation can be moderate.
I2P_LLM_TEMPERATURE_DEFAULT=0.7
I2P_LLM_TEMPERATURE_STORY_GENERATOR=0.7
I2P_LLM_TEMPERATURE_STORY_GENERATOR_REWRITE=0.3
I2P_LLM_TEMPERATURE_STORY_REFLECTOR=0.5
I2P_LLM_TEMPERATURE_PATTERN_SELECTOR=0.3
I2P_LLM_TEMPERATURE_IDEA_FUSION=0.7
I2P_LLM_TEMPERATURE_IDEA_FUSION_STAGE2=0.8
I2P_LLM_TEMPERATURE_IDEA_FUSION_STAGE3=0.9
I2P_LLM_TEMPERATURE_CRITIC_MAIN=0.0
I2P_LLM_TEMPERATURE_CRITIC_REPAIR=0.0
I2P_LLM_TEMPERATURE_IDEA_PACKAGING_PARSE=0.0
I2P_LLM_TEMPERATURE_IDEA_PACKAGING_PATTERN_GUIDED=0.3
I2P_LLM_TEMPERATURE_IDEA_PACKAGING_JUDGE=0.0

# -----------------------------
# Embedding (optional overrides)
# -----------------------------
# If not set, Embedding uses:
# - EMBEDDING_API_URL=https://api.openai.com/v1/embeddings
# - EMBEDDING_MODEL=text-embedding-3-large (example)
# - EMBEDDING_API_KEY falls back to LLM_API_KEY
# Tip: For frequent switching, set I2P_INDEX_DIR_MODE=auto_profile to auto-select
# per-embedding index dirs (no manual profile scripts needed). You can still override
# I2P_NOVELTY_INDEX_DIR / I2P_RECALL_INDEX_DIR if you prefer.
EMBEDDING_API_URL=https://api.openai.com/v1/embeddings
EMBEDDING_MODEL=text-embedding-3-large
EMBEDDING_API_KEY=your_embedding_key_here

# Optional: auto profile index directories
I2P_INDEX_DIR_MODE=auto_profile

# -----------------------------
# Idea Packaging (optional; default off)
# -----------------------------
# Enable idea packaging (pattern-guided, double recall)
# I2P_IDEA_PACKAGING_ENABLE=1
# I2P_IDEA_PACKAGING_TOPN_PATTERNS=5
# I2P_IDEA_PACKAGING_MAX_EXEMPLAR_PAPERS=8
# I2P_IDEA_PACKAGING_CANDIDATE_K=3
# I2P_IDEA_PACKAGING_SELECT_MODE=llm_then_recall  # llm_then_recall|llm_only|recall_only
# I2P_IDEA_PACKAGING_FORCE_EN_QUERY=1

# -----------------------------
# Run logging (repo root `log/`)
# -----------------------------
# 1 = enable structured run logs under `log/run_.../`
# 0 = disable run logs (pipeline still runs)
I2P_ENABLE_LOGGING=1

# Optional: override log output directory (absolute path recommended)
# I2P_LOG_DIR=/abs/path/to/log
# Optional: max chars saved for prompt/response per call (avoid huge JSONL)
# I2P_LOG_MAX_TEXT_CHARS=20000

# -----------------------------
# Results bundling (repo root `results/`)
# -----------------------------
# 1 = enable bundling final artifacts under `results/run_.../`
# 0 = disable bundling (pipeline still runs)
I2P_RESULTS_ENABLE=1

# Results are always copied (no symlinks) to make artifacts portable across platforms.
# Optional: override results output directory (absolute path recommended)
# I2P_RESULTS_DIR=/abs/path/to/results

# -----------------------------
# Critic strictness (quality)
# -----------------------------
# 1 = strict JSON mode (quality-first): critic JSON invalid -> retry -> still invalid => fail the run
# 0 = allow non-strict behavior (useful for offline smoke tests when no API key)
I2P_CRITIC_STRICT_JSON=1

# How many retries after the first failure (default 2)
I2P_CRITIC_JSON_RETRIES=2

# -----------------------------
# Blind Judge (τ calibration)
# -----------------------------
# Path to τ file produced by fit_judge_tau.py (non-secret)
I2P_JUDGE_TAU_PATH=Paper-KG-Pipeline/output/judge_tau.json
#
# Fallback τ values if the τ file is missing (useful for cheap smoke tests)
I2P_TAU_METHODOLOGY=1.0
I2P_TAU_NOVELTY=1.4
I2P_TAU_STORYTELLER=1.0
#
# Coach layer (field-level edit guidance)
I2P_CRITIC_COACH_ENABLE=1
I2P_CRITIC_COACH_TEMPERATURE=0.3
I2P_CRITIC_COACH_MAX_TOKENS=4096

# -----------------------------
# Pass rule (pattern-aware)
# -----------------------------
# Default is the objective "Scheme B":
# - at least 2 of 3 role scores >= pattern q75
# - and avg_score >= pattern q50
#
# If pattern has too few papers (see I2P_PASS_MIN_PATTERN_PAPERS), fallback is controlled by I2P_PASS_FALLBACK.
# I2P_PASS_MODE=two_of_three_q75_and_avg_ge_q50
# I2P_PASS_MIN_PATTERN_PAPERS=20
# I2P_PASS_FALLBACK=global   # global|fixed
# I2P_PASS_SCORE=7.0         # only used when fallback=fixed or distribution unavailable

# -----------------------------
# Advanced: anchors & scoring
# -----------------------------
# Quantiles for initial anchors (comma-separated floats)
I2P_ANCHOR_QUANTILES=0.05,0.15,0.25,0.35,0.5,0.65,0.75,0.85,0.95
I2P_ANCHOR_MAX_INITIAL=11
I2P_ANCHOR_MAX_EXEMPLARS=2
# Optional: densify candidates around score buckets (speed-first)
I2P_ANCHOR_BUCKET_SIZE=1.0
I2P_ANCHOR_BUCKET_COUNT=3
#
# Deterministic score inference grid (smaller = slower but more precise)
# I2P_GRID_STEP=0.01
# Densify triggers when loss/high violations/low avg_strength
I2P_DENSIFY_LOSS_THRESHOLD=0.05
I2P_DENSIFY_MIN_AVG_CONF=0.35
I2P_ANCHOR_DENSIFY_ENABLE=0

# -----------------------------
# Local novelty check (A方案)
# -----------------------------
# Enable local novelty check against nodes_paper.json
I2P_NOVELTY_ENABLE=1
# Do NOT auto-build novelty index during run (quality-first + predictable)
I2P_NOVELTY_AUTO_BUILD_INDEX=1
# Offline build batch size
I2P_NOVELTY_INDEX_BUILD_BATCH_SIZE=32
# Action on high similarity: report_only | pivot | fail
I2P_NOVELTY_ACTION=pivot
# Max pivot attempts when similarity is high
I2P_NOVELTY_MAX_PIVOTS=2

# -----------------------------
# Index auto-prepare (one-command run)
# -----------------------------
# 1 = auto-preflight and build missing indexes; 0 = skip preflight
I2P_INDEX_AUTO_PREPARE=1
# 1 = allow auto-build when missing; 0 = fail and ask for manual build
I2P_INDEX_ALLOW_BUILD=1

# -----------------------------
# Final collision threshold (Phase 4)
# -----------------------------
# 1 = enable final verification (Phase 4), 0 = skip
I2P_VERIFICATION_ENABLE=1
# Recommendation: set between novelty.medium_th and novelty.high_th (e.g. 0.82~0.88)
I2P_COLLISION_THRESHOLD=0.88

# -----------------------------
# Recall audit (persist recall candidates)
# -----------------------------
# 1 = enable recall audit, 0 = disable
I2P_RECALL_AUDIT_ENABLE=1
# Top-N pattern scores per path to persist
I2P_RECALL_AUDIT_TOPN=50
# Recall embedding batch params
I2P_RECALL_EMBED_BATCH_SIZE=32
I2P_RECALL_EMBED_MAX_RETRIES=3
I2P_RECALL_EMBED_SLEEP_SEC=0.5
# Recall offline index (optional)
I2P_RECALL_USE_OFFLINE_INDEX=1

# Optional: canonical subdomain taxonomy for Path2 (reduce duplicates & long-tail)
# When enabled, pipeline auto-detects/builds (if I2P_INDEX_ALLOW_BUILD=1) at:
#   {recall_index_dir}/subdomain_taxonomy.json
I2P_SUBDOMAIN_TAXONOMY_ENABLE=1
I2P_SUBDOMAIN_TAXONOMY_PATH=
I2P_SUBDOMAIN_TAXONOMY_STOPLIST_MODE=drop   # drop