Add LLM token

backend/app/config/settings.py

@@ -133,6 +133,42 @@ class Settings(BaseSettings):
     reranker_api_key: str = Field(default="", description="Reranker API 密钥")
     reranker_top_k: int = Field(default=5, description="精排后保留的最终结果数量")
 
+    # ── HyDE (Hypothetical Document Embeddings) ──────────────────────────────
+    # When enabled, the agentic and standard RAG pipelines generate a short
+    # hypothetical answer before retrieval, then embed that text instead of the
+    # raw query. This closes the vocabulary gap between terse queries and longer
+    # document passages, typically improving recall by 15-30% on vague queries.
+    hyde_enabled: bool = Field(default=True, description="启用 HyDE 查询增强（假设文档嵌入）")
+    hyde_max_tokens: int = Field(default=200, description="HyDE 假设段落最大 token 数")
+    # Use a lightweight model for HyDE to reduce latency and cost.
+    # HyDE only needs a short plausible passage — a fast cheap model is sufficient.
+    # Leave empty to fall back to the main llm_provider / llm_model.
+    hyde_llm_provider: str = Field(default="", description="HyDE 专用 LLM 提供商（空则复用主 LLM）")
+    hyde_llm_model: str = Field(default="", description="HyDE 专用 LLM 模型（空则复用主 LLM）")
+
+    # ── Agentic RAG (P0-1) ───────────────────────────────────────────────────
+    # Controls the multi-step reasoning pipeline exposed at /agent/agentic/stream.
+    agentic_max_sub_queries: int = Field(
+        default=4,
+        description="Agentic 模式最大子查询分解数量（compare / multi_hop 意图触发）",
+    )
+    agentic_grounding_threshold: float = Field(
+        default=0.65,
+        description=(
+            "引文锚定 fast-path 阈值：avg_score > 此值且 chunks ≥ 3 时跳过 LLM grounding check，"
+            "直接判定为充分；降低此值可让更多问题触发 LLM 二次验证。"
+        ),
+    )
+    agentic_intent_max_tokens: int = Field(
+        default=200, description="意图分析步骤 LLM 最大 token 数"
+    )
+    agentic_plan_max_tokens: int = Field(
+        default=400, description="查询分解步骤 LLM 最大 token 数"
+    )
+    agentic_grounding_max_tokens: int = Field(
+        default=250, description="引文锚定步骤 LLM 最大 token 数"
+    )
+
     # Keep configuration setup explicit so runtime behavior is easy to reason about.
     milvus_index_type: str = Field(default="IVF_FLAT", description="Milvus索引类型")
     milvus_nlist: int = Field(default=128, description="Milvus nlist参数")