diff --git a/src/genesis/config.py b/src/genesis/config.py
new file mode 100644
index 0000000..0d53a1f
--- /dev/null
+++ b/src/genesis/config.py
@@ -0,0 +1,206 @@
+from __future__ import annotations
+
+import os
+from pathlib import Path
+from typing import Any
+
+import yaml
+from pydantic import BaseModel, Field
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+SECRET_KEYWORDS = ("key", "secret", "token")
+ENV_PREFIX = "GENESIS_"
+
+
+# ---------- 各 yaml 对应的 pydantic 模型 ----------
+
+class ServerConfig(BaseModel):
+    max_upload_mb: int = 100
+    allowed_extensions: list[str] = Field(
+        default_factory=lambda: [".xlsx", ".xls", ".docx", ".pptx", ".java", ".xml", ".yml"]
+    )
+
+
+class AppConfig(BaseModel):
+    name: str = "genesis"
+    version: str = "0.1.0"
+    timezone: str = "Asia/Tokyo"
+    server: ServerConfig = Field(default_factory=ServerConfig)
+    session: dict[str, Any] = Field(default_factory=lambda: {
+        "sqlite_path": "/data/db/genesis.db",
+        "snapshot_dir": "/data/db/snapshots",
+    })
+    paths: dict[str, Any] = Field(default_factory=lambda: {
+        "user_root": "/data/users",
+        "shared_root": "/data/shared",
+    })
+    task_queue: dict[str, Any] = Field(default_factory=lambda: {
+        "backend": "memory",
+        "redis_url": "",
+        "timeout_sec": 600,
+        "retry_default": 2,
+    })
+
+
+class ModelSpec(BaseModel):
+    provider: str = "deepseek"
+    name: str = "deepseek-chat"
+    temperature: float = 0.2
+    max_tokens: int = 4096
+    timeout_sec: int = 60
+    retry_backoff: list[float] = Field(default_factory=lambda: [1.0, 3.0, 7.0])
+
+
+class InferenceModels(BaseModel):
+    primary: ModelSpec = Field(default_factory=ModelSpec)
+    fallback: ModelSpec = Field(default_factory=lambda: ModelSpec(provider="qwen", name="qwen-max"))
+    vision: ModelSpec = Field(default_factory=lambda: ModelSpec(name="deepseek-vl", timeout_sec=90))
+
+
+class LlmCallsConfig(BaseModel):
+    token_estimation: str = "tiktoken"
+    max_context_tokens: int = 32000
+    truncation_policy: dict[str, Any] = Field(default_factory=lambda: {
+        "priority": ["shrink_rule_chunks", "summarize_history", "truncate_data"],
+    })
+
+
+class StructuredOutputConfig(BaseModel):
+    max_parse_retry: int = 2
+
+
+class PromptRegistryConfig(BaseModel):
+    prompts_dir: str = "./prompts"
+    default_version: str = "latest"
+
+
+class InferenceConfig(BaseModel):
+    models: InferenceModels = Field(default_factory=InferenceModels)
+    llm_calls: LlmCallsConfig = Field(default_factory=LlmCallsConfig)
+    structured_output: StructuredOutputConfig = Field(default_factory=StructuredOutputConfig)
+    prompt_registry: PromptRegistryConfig = Field(default_factory=PromptRegistryConfig)
+
+
+class EmbeddingConfig(BaseModel):
+    model: str = "BAAI/bge-small-zh-v1.5"
+    device: str = "cpu"
+    max_batch_size: int = 32
+    cache_dir: str = "/data/shared/models"
+
+
+class ChromaStoreConfig(BaseModel):
+    persist_dir: str = "/data/shared/rules-handbook/chroma"
+
+
+class QdrantStoreConfig(BaseModel):
+    url: str = "http://qdrant:6333"
+    api_key: str = ""
+
+
+class VectorStoreConfig(BaseModel):
+    adapter: str = "chroma"
+    chroma: ChromaStoreConfig = Field(default_factory=ChromaStoreConfig)
+    qdrant: QdrantStoreConfig = Field(default_factory=QdrantStoreConfig)
+
+
+class ChunkingConfig(BaseModel):
+    word_max_tokens: int = 512
+    excel_rule_block_rows: int = 10
+    ppt_pages_per_chunk: int = 2
+    min_tokens: int = 30
+
+
+class RetrievalConfig(BaseModel):
+    channel_top_k: int = 10
+    rrf_k: int = 60
+    default_top_k: int = 5
+    contextual_enrichment: bool = True
+
+
+class RagConfig(BaseModel):
+    embedding: EmbeddingConfig = Field(default_factory=EmbeddingConfig)
+    vector_store: VectorStoreConfig = Field(default_factory=VectorStoreConfig)
+    chunking: ChunkingConfig = Field(default_factory=ChunkingConfig)
+    retrieval: RetrievalConfig = Field(default_factory=RetrievalConfig)
+
+
+# ---------- 加载辅助 ----------
+
+def _expand_env(data: Any) -> Any:
+    """递归展开 ${VAR} 占位（读环境变量，缺失→空串）"""
+    if isinstance(data, dict):
+        return {k: _expand_env(v) for k, v in data.items()}
+    if isinstance(data, list):
+        return [_expand_env(v) for v in data]
+    if isinstance(data, str) and data.startswith("${") and data.endswith("}"):
+        return os.environ.get(data[2:-1], "")
+    return data
+
+
+def _deep_merge(base: dict, override: dict) -> dict:
+    """递归合并：override 覆盖 base；非 dict 值直接取 override 存在者"""
+    out = dict(base)
+    for k, v in override.items():
+        if isinstance(v, dict) and isinstance(out.get(k), dict):
+            out[k] = _deep_merge(out[k], v)
+        else:
+            out[k] = v
+    return out
+
+
+def _env_overrides() -> dict:
+    """收集 GENESIS_ 前缀的条目为嵌套 dict，__ 为嵌套分隔（键统一小写以匹配 yaml）"""
+    result: dict[str, Any] = {}
+    for key, value in os.environ.items():
+        if key.startswith(ENV_PREFIX):
+            parts = key[len(ENV_PREFIX):].split("__")
+            node = result
+            for part in parts[:-1]:
+                node = node.setdefault(part.lower(), {})
+            node[parts[-1].lower()] = value
+    return result
+
+
+def _load_yaml(config_dir: Path, name: str) -> dict:
+    path = config_dir / f"{name}.yaml"
+    if not path.exists():
+        return {}
+    with path.open("r", encoding="utf-8") as f:
+        return yaml.safe_load(f) or {}
+
+
+def _redact(data: dict) -> dict:
+    out = {}
+    for k, v in data.items():
+        if any(kw in str(k).lower() for kw in SECRET_KEYWORDS):
+            out[k] = "***"
+        elif isinstance(v, dict):
+            out[k] = _redact(v)
+        else:
+            out[k] = v
+    return out
+
+
+# ---------- 根 Settings ----------
+
+class Settings(BaseSettings):
+    model_config = SettingsConfigDict(env_prefix=ENV_PREFIX, env_file=".env", extra="ignore")
+
+    app: AppConfig = Field(default_factory=AppConfig)
+    inference: InferenceConfig = Field(default_factory=InferenceConfig)
+    rag: RagConfig = Field(default_factory=RagConfig)
+
+    @classmethod
+    def from_dir(cls, config_dir: Path | str) -> "Settings":
+        config_dir = Path(config_dir)
+        raw = {
+            "app": _load_yaml(config_dir, "app"),
+            "inference": _load_yaml(config_dir, "inference"),
+            "rag": _load_yaml(config_dir, "rag"),
+        }
+        env = _env_overrides()
+        merged = {k: _deep_merge(raw[k], env.get(k, {})) for k in raw}
+        return cls(**{k: _expand_env(v) for k, v in merged.items()})
+
+    def get_redacted(self) -> dict:
+        return _redact(self.model_dump(mode="json"))
\ No newline at end of file
diff --git a/tests/fixtures/app.yaml b/tests/fixtures/app.yaml
new file mode 100644
index 0000000..e4e5e7a
--- /dev/null
+++ b/tests/fixtures/app.yaml
@@ -0,0 +1,7 @@
+server:
+  max_upload_mb: 10
+session:
+  sqlite_path: "C:/tmp/genesis.db"
+task_queue:
+  backend: memory
+  timeout_sec: 300
\ No newline at end of file
diff --git a/tests/fixtures/inference.yaml b/tests/fixtures/inference.yaml
new file mode 100644
index 0000000..e362ad3
--- /dev/null
+++ b/tests/fixtures/inference.yaml
@@ -0,0 +1,8 @@
+models:
+  primary:
+    provider: deepseek
+    name: deepseek-chat
+llm_calls:
+  max_context_tokens: 16000
+structured_output:
+  max_parse_retry: 3
\ No newline at end of file
diff --git a/tests/fixtures/rag.yaml b/tests/fixtures/rag.yaml
new file mode 100644
index 0000000..4d50bbb
--- /dev/null
+++ b/tests/fixtures/rag.yaml
@@ -0,0 +1,9 @@
+embedding:
+  model: BAAI/bge-small-zh-v1.5
+vector_store:
+  adapter: chroma
+  qdrant:
+    url: http://qdrant:6333
+    api_key: ${QDRANT_API_KEY}
+retrieval:
+  rrf_k: 42
\ No newline at end of file
diff --git a/tests/test_config.py b/tests/test_config.py
new file mode 100644
index 0000000..3f28bc7
--- /dev/null
+++ b/tests/test_config.py
@@ -0,0 +1,51 @@
+from pathlib import Path
+
+from genesis.config import Settings
+
+FIXTURES = Path(__file__).parent / "fixtures"
+
+
+def test_from_dir_maps_yaml_fields():
+    s = Settings.from_dir(FIXTURES)
+    assert s.app.server.max_upload_mb == 10
+    assert s.app.session["sqlite_path"] == "C:/tmp/genesis.db"
+    assert s.app.task_queue["timeout_sec"] == 300
+    assert s.inference.models.primary.name == "deepseek-chat"
+    assert s.inference.llm_calls.max_context_tokens == 16000
+    assert s.inference.structured_output.max_parse_retry == 3
+    assert s.rag.embedding.model == "BAAI/bge-small-zh-v1.5"
+    assert s.rag.retrieval.rrf_k == 42
+
+
+def test_defaults_when_dir_empty(tmp_path):
+    s = Settings.from_dir(tmp_path)
+    assert s.app.name == "genesis"
+    assert s.app.server.max_upload_mb == 100
+    assert s.app.task_queue["backend"] == "memory"
+    assert s.inference.models.primary.name == "deepseek-chat"
+    assert s.rag.embedding.model == "BAAI/bge-small-zh-v1.5"
+    assert s.rag.retrieval.rrf_k == 60
+
+
+def test_env_override_yaml(monkeypatch):
+    monkeypatch.setenv("GENESIS_APP__SERVER__MAX_UPLOAD_MB", "25")
+    s = Settings.from_dir(FIXTURES)
+    assert s.app.server.max_upload_mb == 25
+
+
+def test_env_create_missing_key(monkeypatch):
+    monkeypatch.setenv("GENESIS_RAG__RETRIEVAL__DEFAULT_TOP_K", "7")
+    s = Settings.from_dir(FIXTURES)
+    assert s.rag.retrieval.default_top_k == 7
+
+
+def test_env_placeholder_expansion(monkeypatch):
+    monkeypatch.setenv("QDRANT_API_KEY", "sk-test-xyz")
+    s = Settings.from_dir(FIXTURES)
+    assert s.rag.vector_store.qdrant.api_key == "sk-test-xyz"
+
+
+def test_redacted_hides_secrets():
+    s = Settings.from_dir(FIXTURES)
+    red = s.get_redacted()
+    assert red["rag"]["vector_store"]["qdrant"]["api_key"] == "***"
\ No newline at end of file