运行报错 Novelty check requires embeddings, but embedding is unavailable #27
Description
报错信息如下：
✅ Story终稿已根据Reflection建议生成
🔄 迭代轮次: 2/3
================================================================================
🔍 Phase 3: Multi-Agent Critic (多智能体评审 - Anchored)
📝 Reviewer A (Methodology) 评审中...
评分: 7.1/10
反馈: 【实际输出内容省略】.
📝 Reviewer B (Novelty) 评审中...
评分: 7.1/10
反馈: 【实际输出内容省略】.
📝 Reviewer C (Storyteller) 评审中...
评分: 6.9/10
反馈: 【实际输出内容省略】.
📊 诊断信息:
分数分布: [7.099999999999892, 7.109999999999892, 6.909999999999896]
最低分评审员: Reviewer C (Storyteller), 分数: 6.909999999999896
📊 评审结果: 平均分 7.04/10 - ✅ PASS
🏆 更新全局最佳版本: 得分 7.04 (迭代 2)
✅ 评审通过，进入查重验证阶段
❌ 错误: Novelty check requires embeddings, but embedding is unavailable
Traceback (most recent call last):
File "F:\software\ChatGPT\Idea2Paper-main\Paper-KG-Pipeline\scripts\idea2story_pipeline.py", line 304, in main
result = pipeline.run()
File "F:\software\ChatGPT\Idea2Paper-main\Paper-KG-Pipeline\src\idea2paper\application\pipeline\manager.py", line 512, in run
raise RuntimeError("Novelty check requires embeddings, but embedding is unavailable")
RuntimeError: Novelty check requires embeddings, but embedding is unavailable
环境配置：已经将 Hugging Face 上 paper-embedding 中的两个文件夹放入 paper-KG-Pipeline/output 下
.env.example 文件内容如下：
LLM_API_URL=https://api.siliconflow.cn/v1/chat/completions
LLM_MODEL=Pro/zai-org/GLM-4.7
-----------------------------
Embedding (optional overrides)
-----------------------------
If not set, Embedding uses:
- EMBEDDING_API_URL=https://api.siliconflow.cn/v1/embeddings
- EMBEDDING_MODEL=Qwen/Qwen3-Embedding-8B
- EMBEDDING_API_KEY falls back to SILICONFLOW_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.siliconflow.cn/v1/embeddings
EMBEDDING_MODEL=Qwen/Qwen3-Embedding-8B
EMBEDDING_API_KEY=your_embedding_key_here
Optional: auto profile index directories
I2P_INDEX_DIR_MODE=auto_profile
-----------------------------
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
Bundling mode: link (preferred) or copy
- link: create symlink if possible, fallback to copy
- copy: always duplicate files
I2P_RESULTS_MODE=link
-----------------------------
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
-----------------------------
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 the 5 fixed anchors (comma-separated floats)
I2P_ANCHOR_QUANTILES=0.1,0.25,0.5,0.75,0.9
I2P_ANCHOR_MAX_INITIAL=7
I2P_ANCHOR_MAX_TOTAL=9
I2P_ANCHOR_MAX_EXEMPLARS=2
I2P_DENSIFY_OFFSETS=-0.5,0.5,-0.25,0.25
I2P_SIGMOID_K=1.2
I2P_GRID_STEP=0.01
I2P_DENSIFY_LOSS_THRESHOLD=0.03
I2P_DENSIFY_MIN_AVG_CONF=0.45
I2P_ANCHOR_DENSIFY_ENABLE=0 # disable adaptive densify to reduce latency
-----------------------------
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
i2p_config.json文件，没有修改任何内容