RAG系统构建完全指南：打造企业级知识问答系统

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266

from typing import List, Dict, Any
from sentence_transformers import SentenceTransformer
import numpy as np

class EnterpriseRAG:
    """企业级RAG系统"""

    def __init__(self, config: Dict[str, Any]):
        # 配置
        self.config = config

        # 嵌入模型
        self.embedder = SentenceTransformer(
            config.get("embed_model", "sentence-transformers/all-MiniLM-L6-v2")
        )

        # 向量数据库
        self.vector_db = self._init_vector_db(config.get("vector_db"))

        # LLM客户端
        self.llm = self._init_llm(config.get("llm"))

    def _init_vector_db(self, db_config: Dict):
        """初始化向量数据库"""

        db_type = db_config.get("type", "faiss")

        if db_type == "faiss":
            import faiss
            dimension = self.embedder.get_sentence_embedding_dimension()
            index = faiss.IndexFlatIP(dimension)
            return {"type": "faiss", "index": index}

        elif db_type == "pinecone":
            import pinecone
            pinecone.init(
                api_key=db_config["api_key"],
                environment=db_config["environment"]
            )
            return {"type": "pinecone"}

        elif db_type == "milvus":
            from pymilvus import connections, Collection
            connections.connect(
                alias="default",
                host=db_config["host"],
                port=db_config["port"]
            )
            return {"type": "milvus"}

    def _init_llm(self, llm_config: Dict):
        """初始化LLM"""

        provider = llm_config.get("provider", "anthropic")

        if provider == "anthropic":
            from anthropic import Anthropic
            return Anthropic(api_key=llm_config["api_key"])

        elif provider == "openai":
            from openai import OpenAI
            return OpenAI(api_key=llm_config["api_key"])

        elif provider == "local":
            # 使用本地模型
            return None

    def index_documents(self, documents: List[Dict[str, Any]]):
        """索引文档"""

        # 提取文本
        texts = [doc["content"] for doc in documents]

        # 文本切分
        chunks = self._chunk_texts(texts, chunk_size=500, overlap=50)

        # 生成嵌入
        embeddings = self.embedder.encode(
            [chunk["text"] for chunk in chunks]
        )

        # 存储到向量数据库
        self._store_embeddings(chunks, embeddings)

    def _chunk_texts(
        self,
        texts: List[str],
        chunk_size: int = 500,
        overlap: int = 50
    ) -> List[Dict]:
        """文本切分"""

        chunks = []

        for idx, text in enumerate(texts):
            # 按段落切分
            paragraphs = text.split("\n\n")

            current_chunk = ""
            chunk_id = 0

            for para in paragraphs:
                if len(current_chunk) + len(para) > chunk_size:
                    if current_chunk:
                        chunks.append({
                            "text": current_chunk.strip(),
                            "doc_id": idx,
                            "chunk_id": chunk_id
                        })
                        chunk_id += 1
                    current_chunk = para[overlap:]
                else:
                    current_chunk += "\n\n" + para

            if current_chunk:
                chunks.append({
                    "text": current_chunk.strip(),
                    "doc_id": idx,
                    "chunk_id": chunk_id
                })

        return chunks

    def _store_embeddings(self, chunks: List[Dict], embeddings: np.ndarray):
        """存储嵌入"""

        db_type = self.vector_db["type"]

        if db_type == "faiss":
            import faiss

            # 归一化
            faiss.normalize_L2(embeddings)

            # 添加到索引
            self.vector_db["index"].add(
                embeddings.astype("float32")
            )

            # 存储元数据
            if "metadata" not in self.vector_db:
                self.vector_db["metadata"] = []
            self.vector_db["metadata"].extend(chunks)

        elif db_type == "pinecone":
            import pinecone

            index = pinecone.Index(self.vector_db["index_name"])

            for i, (chunk, embedding) in enumerate(zip(chunks, embeddings)):
                index.upsert([
                    (str(i), embedding.tolist(), {
                        "text": chunk["text"],
                        "doc_id": chunk["doc_id"]
                    })
                ])

    def retrieve(self, query: str, top_k: int = 5) -> List[Dict]:
        """检索相关文档"""

        # 生成查询嵌入
        query_embedding = self.embedder.encode([query])

        db_type = self.vector_db["type"]

        if db_type == "faiss":
            import faiss

            # 归一化
            faiss.normalize_L2(query_embedding)

            # 搜索
            scores, indices = self.vector_db["index"].search(
                query_embedding.astype("float32"),
                top_k
            )

            # 组装结果
            results = []
            for score, idx in zip(scores[0], indices[0]):
                if idx < len(self.vector_db["metadata"]):
                    meta = self.vector_db["metadata"][idx]
                    results.append({
                        "text": meta["text"],
                        "score": float(score),
                        "doc_id": meta["doc_id"]
                    })

            return results

        elif db_type == "pinecone":
            import pinecone

            index = pinecone.Index(self.vector_db["index_name"])

            query_result = index.query(
                vector=query_embedding[0].tolist(),
                top_k=top_k,
                include_metadata=True
            )

            return [
                {
                    "text": match["metadata"]["text"],
                    "score": match["score"],
                    "doc_id": match["metadata"]["doc_id"]
                }
                for match in query_result["matches"]
            ]

    def generate(self, query: str, context: List[Dict]) -> str:
        """生成回答"""

        # 构建提示
        context_text = "\n\n".join([
            f"[文档{i+1}] {doc['text']}"
            for i, doc in enumerate(context)
        ])

        prompt = f"""
基于以下文档回答问题。如果文档中没有相关信息，请明确说明。

{context_text}

问题：{query}

回答：
"""

        # 调用LLM
        if self.llm:
            response = self.llm.messages.create(
                model="claude-3-5-sonnet-20241022",
                max_tokens=2048,
                messages=[{"role": "user", "content": prompt}]
            )
            return response.content[0].text
        else:
            # 本地模型
            pass

    def ask(self, query: str, top_k: int = 5) -> Dict[str, Any]:
        """问答接口"""

        # 检索
        retrieved_docs = self.retrieve(query, top_k=top_k)

        # 检查相关性
        if retrieved_docs and retrieved_docs[0]["score"] < 0.5:
            return {
                "answer": "抱歉，我在知识库中没有找到相关信息。",
                "sources": [],
                "confidence": "low"
            }

        # 生成
        answer = self.generate(query, retrieved_docs)

        return {
            "answer": answer,
            "sources": [
                {"doc_id": doc["doc_id"], "score": doc["score"]}
                for doc in retrieved_docs
            ],
            "confidence": "high" if retrieved_docs[0]["score"] > 0.7 else "medium"
        }

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53

class HybridRetriever:
    """混合检索：向量+关键词"""

    def __init__(self, vector_db, keyword_index):
        self.vector_db = vector_db
        self.keyword_index = keyword_index

    def retrieve(self, query: str, alpha: float = 0.5, k: int = 5):
        """
        混合检索
        alpha: 向量检索权重 (0-1)
        1-alpha: 关键词检索权重
        """

        # 向量检索
        vector_results = self.vector_db.search(query, k=k*2)

        # 关键词检索
        keyword_results = self.keyword_index.search(query, k=k*2)

        # 结果融合（RRF算法）
        fused = self.reciprocal_rank_fusion(
            {
                "vector": vector_results,
                "keyword": keyword_results
            },
            weights={"vector": alpha, "keyword": 1-alpha}
        )

        return sorted(
            fused.items(),
            key=lambda x: x[1],
            reverse=True
        )[:k]

    def reciprocal_rank_fusion(self, results: Dict, weights: Dict, k: int = 60):
        """RRF融合算法"""

        fused = {}

        for method, items in results.items():
            weight = weights.get(method, 1.0)

            for rank, item in enumerate(items):
                item_id = item["id"]

                if item_id not in fused:
                    fused[item_id] = 0

                # RRF公式
                fused[item_id] += weight / (k + rank + 1)

        return fused

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

class Reranker:
    """重排序器"""

    def __init__(self):
        from sentence_transformers import CrossEncoder
        self.reranker = CrossEncoder(
            "cross-encoder/ms-marco-MiniLM-L-6-v2"
        )

    def rerank(self, query: str, documents: List[Dict], top_k: int = 5):
        """重排序"""

        # 计算查询-文档相关性分数
        pairs = [[query, doc["text"]] for doc in documents]
        scores = self.reranker.predict(pairs)

        # 更新分数
        for doc, score in zip(documents, scores):
            doc["rerank_score"] = float(score)

        # 按新分数排序
        reranked = sorted(
            documents,
            key=lambda x: x["rerank_score"],
            reverse=True
        )

        return reranked[:top_k]

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93

class RAGEvaluator:
    """RAG系统评估"""

    def evaluate(
        self,
        test_queries: List[Dict],
        rag_system: EnterpriseRAG
    ) -> Dict[str, float]:
        """
        test_queries: [
            {
                "query": "什么是机器学习？",
                "expected_answer": "机器学习是...",
                "expected_docs": [1, 3, 5]
            }
        ]
        """

        metrics = {
            "retrieval_precision": [],
            "retrieval_recall": [],
            "answer_relevance": [],
            "answer_faithfulness": []
        }

        for test in test_queries:
            # 执行查询
            result = rag_system.ask(test["query"])

            # 检索评估
            retrieved_doc_ids = [s["doc_id"] for s in result["sources"]]
            expected_doc_ids = test["expected_docs"]

            # 精确率
            precision = len(
                set(retrieved_doc_ids) & set(expected_doc_ids)
            ) / len(retrieved_doc_ids)
            metrics["retrieval_precision"].append(precision)

            # 召回率
            recall = len(
                set(retrieved_doc_ids) & set(expected_doc_ids)
            ) / len(expected_doc_ids)
            metrics["retrieval_recall"].append(recall)

            # 答案相关性
            relevance = self._compute_relevance(
                result["answer"],
                test["expected_answer"]
            )
            metrics["answer_relevance"].append(relevance)

            # 忠实度
            faithfulness = self._compute_faithfulness(
                result["answer"],
                [s["text"] for s in result["sources"]]
            )
            metrics["answer_faithfulness"].append(faithfulness)

        # 计算平均指标
        return {
            k: sum(v) / len(v)
            for k, v in metrics.items()
        }

    def _compute_relevance(self, generated: str, expected: str) -> float:
        """计算相关性（使用余弦相似度）"""

        from sentence_transformers import SentenceTransformer
        embedding_model = SentenceTransformer("all-MiniLM-L6-v2")

        emb1 = embedding_model.encode([generated])
        emb2 = embedding_model.encode([expected])

        from sklearn.metrics.pairwise import cosine_similarity
        return cosine_similarity(emb1, emb2)[0][0]

    def _compute_faithfulness(self, answer: str, contexts: List[str]) -> float:
        """计算忠实度（答案是否基于上下文）"""

        # 简单方法：检查答案中的关键实体是否在上下文中
        import spacy
        nlp = spacy.load("zh_core_web_sm")

        doc = nlp(answer)
        entities = [ent.text for ent in doc.ents]

        found = sum(
            1 for ent in entities
            if any(ent in ctx for ctx in contexts)
        )

        return found / len(entities) if entities else 1.0

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46

from fastapi import FastAPI
from pydantic import BaseModel

app = FastAPI(title="Enterprise RAG API")

class Question(BaseModel):
    text: str
    top_k: int = 5

class Answer(BaseModel):
    answer: str
    sources: List[Dict]
    confidence: str

# 初始化RAG系统
rag = EnterpriseRAG({
    "embed_model": "sentence-transformers/all-MiniLM-L6-v2",
    "vector_db": {
        "type": "pinecone",
        "api_key": "your-api-key",
        "environment": "us-east-1-aws",
        "index_name": "enterprise-kb"
    },
    "llm": {
        "provider": "anthropic",
        "api_key": "your-api-key"
    }
})

# 启动时索引文档
@app.on_event("startup")
async def startup():
    documents = load_documents_from_s3()
    rag.index_documents(documents)

@app.post("/api/ask", response_model=Answer)
async def ask(question: Question):
    """问答接口"""

    result = rag.ask(question.text, question.top_k)

    return Answer(**result)

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)