实用笔记：第三部分：高级RAG技术——HyDE、自我查询与智能代理工作流
《实用笔记》操作指南：第三部分——专家级RAG技术：HyDE、自我查询与代理工作流，以及为正在开发RAG系统的团队提供的合同模板、检查清单及可直接插入的代码片段。
本指南将逐步构建从原材料到可运行系统的完整流程，内容涵盖：第三部分：高级RAG技术——HyDE、自我查询与智能工作流。重点在于可操作的步骤、明确的检查点，以及可直接放入代码库的代码，无需猜测其用途。 在修改代码之前，应先明确输入参数、各步骤的负责人以及完成标准。操作人员应能够从已知的检查点重新运行相应步骤，而无需推测隐藏状态。 除了功能结果外，还需记录执行时间以及Token或查询成本。提前了解成本情况，可避免在流程从演示环境过渡到共享环境时出现意外费用。
1. HyDE：假设性文档嵌入技术
在实现1. HyDE：假设性文档嵌入时，首先明确相关约定：所需输入、成功标志以及部分失败时的处理方式。这样的清单能确保后续的代码修改保持一致性。 将配置信息置于应用程序代码之外。环境文件、密钥存储以及功能开关应集中存放，以便操作人员无需查看整个系统结构即可进行审计。 在调整提示词之前，先使用固定的问题集来衡量召回率。仅仅更换提示词往往无法改善较差的检索效果。
核心见解
在处理《洞察》相关内容时，首先写下契约：所需的输入参数、成功信号以及部分失败时的处理方式。这样的清单能确保后续的代码修改保持一致性。 同时记录正常流程和异常恢复流程。重试机制、人工审核环节以及死信处理都是产品本身的组成部分，而非后续需要优化的内容。 在调整提示词之前，先使用固定的问题集来衡量召回率。仅仅更换提示词很难解决检索效果不佳的问题。
HyDE的工作原理
在研究 How HyDE Works 时，首先需列出相关约定：所需的输入参数、成功信号以及部分失败时的处理方式。这样的清单能确保后续的代码修改保持一致性。 相比冗长的脚本，应优先选择小型且可测试的单元。当某个步骤出现故障时，故障点应指向单一责任模块，而非复杂的流程链。 在调整提示词之前，先使用固定的问题集来测试召回率。仅仅更换提示词很难解决检索效果不佳的问题。
Step 1: User Query
"What are ISRO's major achievements?"
Step 2: Generate Hypothetical Answer (using LLM)
"ISRO has achieved several milestones including reaching Mars orbit
in 2014, successfully landing Chandrayaan-3 on the Moon's south pole
in 2023, and launching satellites for multiple countries at low cost..."
Step 3: Embed the Hypothetical Answer
[0.23, -0.45, 0.67, ...] ← This lives in document space!
Step 4: Retrieve Documents Similar to Hypothetical Answer
Now we're comparing document-to-document, not query-to-document
Step 5: Generate Final Answer
Use retrieved docs + original query → Better answer
HyDE 为何有效
在研究“Why HyDE Works”时，首先需列出相关契约：所需输入、成功信号以及部分失败时的处理方式。这样的清单能确保后续的代码修改保持一致性。 将此阶段视为输入与验证后输出之间的契约。为相关成果命名，明确成功判定标准，并杜绝无声的半完成状态。 在调整提示词之前，先使用固定的问题集来衡量召回率。仅仅更换提示词很难改善较差的检索效果。
何时使用 HyDE
在研究何时使用HyDE时，首先需列出相关约定：所需输入、成功信号以及部分失败时的处理方式。这样的清单能确保后续的代码修改保持一致性。 在功能结果旁记录执行时间以及令牌或查询成本。提前了解成本情况，可避免从演示环境过渡到共享环境时出现意外费用。 在调整提示词之前，先用固定的问题集测试召回率。仅仅更换提示词往往无法改善较差的检索效果。
HyDE与传统RAG对比：实际案例
在研究 HyDE 与传统 RAG：实际案例时，首先需明确规范：所需输入、成功标志以及部分失败时的处理方式。这样的检查清单能确保后续的代码修改保持一致性。 应将配置信息置于应用程序代码之外。环境文件、密钥存储和功能开关应集中存放，以便操作人员无需查看整个系统结构即可进行审计。 在调整提示词之前，先使用固定的问题集来衡量召回率。仅仅更换提示词很难解决检索效果不佳的问题。
2. 上下文压缩：精度优于数量
在学习“2. 上下文压缩：精度优于数量”这一内容时，首先需明确相关约定：所需的输入参数、成功信号以及部分失败时的处理方式。这样的清单能确保后续的代码修改保持一致性。 同时记录正常流程与异常恢复流程。重试机制、人工审核环节以及错误消息处理都是产品本身的一部分，而非后续需要补充的功能。 在调整提示词之前，应先使用固定的问题集来衡量召回率。仅仅更换提示词往往无法解决检索效果不佳的问题。
问题所在
在解决相关问题时，首先需明确规范：所需的输入参数、成功标志以及部分失败时的处理方式。这份清单能确保后续的代码修改保持一致性。 相较于冗长的脚本，应优先选择小型且可测试的单元。当某一步骤失败时，故障应指向单一责任模块，而非复杂的流程链。 在调整提示词之前，先使用固定的问题集测试召回率。仅仅更换提示词很难改善较差的信息检索效果。 在解决相关问题时，首先需明确规范：所需的输入参数、成功标志以及部分失败时的处理方式。这份清单能确保后续的代码修改保持一致性。 在功能结果旁记录执行时间以及token或查询成本。提前了解成本情况，可避免从演示环境过渡到共享环境时出现意外费用。
India's capital city is New Delhi, which is located in the northern
part of the country. The city serves as the center of the Government
of India and houses important governmental buildings including the
Parliament House, Rashtrapati Bhavan, and various ministry buildings.
New Delhi was inaugurated in 1931 and became the capital of India
after independence in 1947. The previous capital was Calcutta, now
known as Kolkata. The decision to move the capital was made by the
British colonial government in 1911...
India's capital city is New Delhi
上下文压缩解决方案
将上下文压缩方案视为可度量的对象时，其效果最佳。在扩大范围之前，先记录一份理想的转录文本、一个故障案例以及回滚说明。 将配置置于应用程序代码之外。环境文件、密钥存储和功能标志应集中存放于一处，以便操作人员无需查看整个系统结构即可进行审计。 将分块策略与检索策略分开。当质量指标发生变化时，修改其中一项不应迫使重新编写另一项。
Query + Chunk → Compressor LLM → Relevant Sentences Only
实施策略
将实施策略视为可度量的指标时，其效果最佳。在扩大范围之前，先记录一个成功的案例、一个失败案例以及回滚说明。同时记录正常流程和恢复流程。重试机制、人工审核环节以及死信处理都是产品本身的一部分，而非后续需要补充的内容。应将分块策略与检索策略分开，当质量指标发生变化时，修改其中一项不应迫使重新编写另一项。
# Traditional: Send full chunks
context = chunk1 + chunk2 + chunk3 # 1500 tokens
# Compressed: Extract relevant parts
for chunk in chunks:
compressed = compressor.extract_relevant(query, chunk)
context.append(compressed) # 300 tokens total
优势
将 Benefits 视为可度量的对象时，其效果最佳。在扩大范围之前，先记录一份理想的输出样本、一个故障案例以及回滚说明。 优先选择小型且可测试的单元，而非庞大的脚本。当某个步骤出现故障时，故障应指向单一责任模块，而非复杂的流程链。 将分块策略与检索策略分开。当质量指标发生变化时，修改其中一项不应迫使重新编写另一项。 将 Benefits 视为可度量的对象时，其效果最佳。在扩大范围之前，先记录一份理想的输出样本、一个故障案例以及回滚说明。 在功能结果旁记录执行时间以及令牌或查询成本。提前了解成本情况，可避免在从演示环境过渡到共享环境时出现意外费用。
3. 自主查询：让大语言模型来决策
第三点：自我查询。在修改代码之前，先确定输入内容、该步骤的负责人以及终止标准。操作人员应能够从已知的检查点重新运行该步骤，而无需猜测隐藏状态。配置信息应置于应用程序代码之外，环境文件、密钥存储和功能标志应集中存放于一个位置，以便操作人员无需查看整个流程即可进行审计。当下一步是代码编写或工具调用时，优先使用具有架构验证的结构化输出，而非自由形式的文字描述。
概念
在修改代码之前，需先明确概念中的输入参数、该步骤的负责人以及终止标准。操作人员应能够从已知的检查点重新运行该步骤，而无需猜测隐藏状态。 需同时记录正常流程与异常恢复流程。重试机制、人工审核环节以及错误处理都是产品本身的组成部分，而非后续需要补充的内容。 必须引用那些真正作为答案依据的段落。如果没有引用，操作人员就无法区分是幻觉内容还是索引缺失导致的错误。
传统方式与自查询方式
在采用传统模式与自查询模式时，应在修改代码之前明确输入参数、该步骤的负责人以及终止标准。操作人员应能够从已知的检查点重新运行该步骤，而无需猜测隐藏状态。 相较于冗长的脚本，更应优先使用小型且可测试的单元。当某个步骤失败时，故障原因应能明确指向单一责任方，而非复杂的流程链。 需引用实际作为答案依据的段落。没有引用的话，操作人员就无法区分是幻觉内容还是索引缺失导致的错误。 在采用传统模式与自查询模式时，应在修改代码之前明确输入参数、该步骤的负责人以及终止标准。操作人员应能够从已知的检查点重新运行该步骤，而无需猜测隐藏状态。 除了功能结果外，还需记录执行时间以及令牌或查询成本。提前了解成本情况，可避免在从演示环境过渡到共享环境时出现意外费用。
# Developer hardcodes everything
query = "Find recent cricket matches"
top_k = 5
filters = {"category": "sports"}
results = retriever.run(query, top_k, filters)
# LLM decides everything
query = "Find recent cricket matches"
# LLM analyzes and decides:
# - Extract metadata: {"sport": "cricket", "recency": "2024"}
# - Set top_k: 10 (wants comprehensive results)
# - Use hybrid search (keyword "matches" + semantic)
results = self_querying_retriever.run(query)
为何自我查询很重要
在研究“为何自我查询很重要”这一主题时，首先需明确相关规范：所需的输入参数、成功标志以及部分失败时的处理方式。这样的检查清单能确保后续的代码修改保持一致性。 应将配置信息置于应用程序代码之外。环境文件、密钥存储以及功能开关应集中存放于一个位置，这样操作人员无需查看整个系统结构即可进行审计。 在调整提示词之前，需先使用固定的问题集来测试信息的召回率。仅仅更换提示词往往无法解决检索效果不佳的问题。
示例：自我查询的实际应用
在完成“实战中的自我查询”示例时，首先列出相关约定：所需输入、成功信号以及部分失败时的处理方式。这样的检查清单能确保后续的代码修改保持一致性。 同时记录正常流程和异常恢复流程。重试机制、人工审核环节以及死信处理都是产品本身的组成部分，而非后续的优化内容。 在调整提示词之前，先使用固定的问题集来衡量检索效果。仅仅更换提示词很难解决检索能力薄弱的问题。
{
"semantic_query": "articles about India",
"metadata_filters": {
"year": 2023,
"content_type": "article"
},
"search_type": "hybrid",
"top_k": 10
}
4. 智能代理型RAG：终极进化
在编写“4. 主动式RAG：终极进化”这部分内容时，首先需明确相关约定：所需的输入参数、成功标志，以及部分失败时的处理方式。这样的清单能确保后续的代码修改保持一致性。 建议采用小型、可测试的单元，而非庞大的脚本。当某个步骤出现故障时，故障应指向单一责任模块，而非复杂的流程链。 在调整提示词之前，先使用固定的问题集来测试召回率。仅仅更换提示词很难解决检索效果不佳的问题。
什么让RAG具备“主动性”？
在研究“是什么让RAG具备‘代理’能力？”时，首先需明确相关约定：所需的输入参数、成功标志，以及部分失败时的处理方式。这样的清单能确保后续的代码修改始终符合初始要求。 将此阶段视为输入与经过验证的输出之间的契约。为相关成果命名，定义成功判定标准，杜绝无声的半完成状态。 在调整提示词之前，先使用固定的问题集来测试召回率。仅仅更换提示词往往无法改善较差的信息检索效果。
代理型RAG架构
在研究代理式RAG架构时，首先需明确相关规范：所需的输入参数、成功标志，以及部分失败时的处理方式。这样的清单能确保后续的代码修改始终符合预期。 在记录功能结果的同时，还需标注处理时间以及令牌或查询成本。提前了解成本情况，可避免在从演示环境过渡到共享环境时出现意外费用。 在调整提示词之前，应先使用固定的问题集来测试召回率。仅仅更换提示词往往无法改善较差的检索效果。
┌─────────────┐
│ User Query │
└──────┬──────┘
│
┌──────▼──────┐
│ Agent (LLM) │ ← Makes decisions
└──────┬──────┘
│
┌──────────────────┼──────────────────┐
│ │ │
┌────▼────┐ ┌─────▼─────┐ ┌─────▼──────┐
│ Tool 1 │ │ Tool 2 │ │ Tool 3 │
│ (RAG) │ │(Web Search│ │(Calculator)│
└────┬────┘ └─────┬─────┘ └─────┬──────┘
│ │ │
└──────────────────┼──────────────────┘
│
┌──────▼──────┐
│ Agent │ ← Synthesizes
└──────┬──────┘
│
┌──────▼──────┐
│ Final │
│ Answer │
└─────────────┘
代理决策流程
在处理智能体决策流程时，首先写下相关规范：所需的输入参数、成功标志以及部分失败时的处理方式。这样的检查清单能确保后续的代码修改保持一致性。 将配置信息置于应用程序代码之外。环境文件、密钥存储以及功能开关应集中存放于一个位置，这样操作人员无需查看整个系统结构即可进行审计。 在调整提示词之前，先使用固定的问题集来衡量检索的覆盖率。仅仅更换提示词往往无法解决检索效果不佳的问题。
Step 1: Analyze Query
- Requires comparison between two countries
- Need current statistics
- Two separate pieces of info needed
Step 2: Plan Actions
- Tool 1: Search knowledge base for India population
- Tool 2: Search knowledge base for China population
- Tool 3: If not found, use web search as fallback
Step 3: Execute
→ Search knowledge base for India: Found
→ Search knowledge base for China: Not found
→ Fallback to web search for China: Found
Step 4: Synthesize
Combine information from both sources into coherent answer
智能体RAG功能
智能体工作流示例
User: "Tell me about ISRO"
Agent: [Uses RAG tool] → Provides answer from knowledge base
User: "What about NASA?"
Agent: [Uses RAG tool] → Not found in knowledge base
[Falls back to web search] → Retrieves info from web
→ Provides answer with source attribution
User: "Compare their budgets"
Agent: [Analyzes] → Needs both ISRO and NASA budget data
[Retrieves from both sources]
[Uses calculator tool for comparison]
→ Provides detailed comparison
智能体RAG的优势
整合所有技术：终极RAG系统
User Query: "What are recent achievements in India's space program?"
↓
1. Self-Querying
LLM analyzes: Needs recent info, space domain
Filters: {topic: "space", recency: "2023-2024"}
↓
2. HyDE Generation
"ISRO achieved remarkable milestones in 2023-2024, including
successful Moon landings and satellite launches..."
↓
3. Hybrid Retrieval + HyDE
Retrieve using both: original query + hypothetical answer
Get top 20 documents
↓
4. Contextual Compression
Extract only sentences about recent achievements
Reduce 20 chunks (10k tokens) → 5 compressed chunks (2k tokens)
↓
5. Agentic Decision
Agent: "Retrieved info looks good, but let me verify with web search"
→ Quick web search for latest news
→ Combines both sources
↓
6. Final Answer
Comprehensive, accurate, up-to-date response with source attribution
何时使用每种技术
HyDE
上下文压缩
自我查询
智能体RAG
逐步代码实现
import os
from pathlib import Path
import json
from typing import List, Dict, Any, Optional
from haystack import Pipeline, Document, component
from haystack.document_stores.in_memory import InMemoryDocumentStore
from haystack.components.retrievers.in_memory import (
InMemoryBM25Retriever,
InMemoryEmbeddingRetriever
)
from haystack.components.embedders import (
SentenceTransformersTextEmbedder,
SentenceTransformersDocumentEmbedder
)
from haystack.components.writers import DocumentWriter
from haystack.components.preprocessors import DocumentSplitter, DocumentCleaner
from haystack.components.converters import TextFileToDocument
from haystack.components.builders.prompt_builder import PromptBuilder
from haystack.components.generators import OpenAIGenerator
from haystack.components.joiners import DocumentJoiner
from haystack.components.rankers import SentenceTransformersSimilarityRanker
from haystack.components.routers import ConditionalRouter
from haystack.utils import Secret
from haystack.dataclasses import ChatMessage
# ============================================================================
# CONFIGURATION
# ============================================================================
# GROQ_API_KEY = "your-groq-api-key-here"
os.environ["GROQ_API_KEY"] = "gsk_!!!"
# Model configurations
EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
RERANKER_MODEL = "cross-encoder/ms-marco-MiniLM-L-6-v2"
GROQ_MODEL = "llama-3.3-70b-versatile"
# Retrieval parameters
BM25_TOP_K = 10
EMBEDDING_TOP_K = 10
RERANKER_TOP_K = 5
# ============================================================================
# READING DATA FILE
# ============================================================================
def load_documents_with_metadata(file_path: str) -> list[Document]:
documents = []
with open(file_path, "r", encoding="utf-8") as f:
for line in f:
data = json.loads(line)
documents.append(
Document(
content=data["content"],
meta=data.get("meta", {})
)
)
return documents
# ============================================================================
# COMPONENT 1: HyDE (HYPOTHETICAL DOCUMENT EMBEDDINGS)
# ============================================================================
@component
class HyDEGenerator:
"""Generate hypothetical documents for HyDE retrieval"""
def __init__(self, llm: OpenAIGenerator):
self.llm = llm
self.template = """Generate a detailed paragraph that would perfectly answer
the following question. Write as if you're providing the ideal answer from a knowledge base.
Question: {query}
Ideal Answer Paragraph:"""
@component.output_types(hypothetical_doc=str)
def run(self, query: str) -> Dict[str, str]:
"""Generate hypothetical document"""
prompt = self.template.format(query=query)
result = self.llm.run(prompt=prompt)
hypothetical_doc = result["replies"][0]
return {"hypothetical_doc": hypothetical_doc}
# ============================================================================
# COMPONENT 2: CONTEXTUAL COMPRESSOR
# ============================================================================
@component
class ContextualCompressor:
"""Extract only relevant sentences from retrieved documents"""
def __init__(self, llm: OpenAIGenerator):
self.llm = llm
self.template = """Given the following document chunk and query, extract ONLY
the sentences that are directly relevant to answering the query.
Return only the relevant sentences, nothing else.
Query: {query}
Document Chunk:
{chunk}
Relevant Sentences:"""
@component.output_types(compressed_documents=List[Document])
def run(self, query: str, documents: List[Document]) -> Dict[str, List[Document]]:
"""Compress documents by extracting relevant content"""
compressed_docs = []
for doc in documents:
prompt = self.template.format(query=query, chunk=doc.content)
result = self.llm.run(prompt=prompt)
compressed_content = result["replies"][0]
# Create new document with compressed content
compressed_doc = Document(
content=compressed_content,
meta=doc.meta,
score=doc.score if hasattr(doc, 'score') else None
)
compressed_docs.append(compressed_doc)
return {"compressed_documents": compressed_docs}
# ============================================================================
# COMPONENT 3: SELF-QUERYING ANALYZER
# ============================================================================
@component
class SelfQueryAnalyzer:
"""Analyze query and extract metadata filters automatically"""
def __init__(self, llm: OpenAIGenerator):
self.llm = llm
self.template = """Analyze the following query and extract:
1. The core semantic query (cleaned, focused version)
2. Any metadata filters that should be applied
Available metadata fields:
- category: geography, politics, economy, sports, science, culture
- year: any year (e.g., 2023, 2024)
- topic: overview, capital, gdp, cricket, space, entertainment, language
- source: any source type
Query: {query}
Respond in this exact format:
SEMANTIC_QUERY: [your semantic query here]
FILTERS: category=value,year=value (or FILTERS: none if no filters apply)"""
@component.output_types(semantic_query=str, filters=Dict[str, Any])
def run(self, query: str) -> Dict[str, Any]:
"""Analyze query and extract filters"""
prompt = self.template.format(query=query)
result = self.llm.run(prompt=prompt)
response = result["replies"][0]
# Parse response
lines = response.strip().split('\n')
semantic_query = query # default
filters = {}
for line in lines:
if line.startswith("SEMANTIC_QUERY:"):
semantic_query = line.replace("SEMANTIC_QUERY:", "").strip()
elif line.startswith("FILTERS:"):
filters_str = line.replace("FILTERS:", "").strip()
if filters_str.lower() != "none":
# Parse filters
for filter_pair in filters_str.split(','):
if '=' in filter_pair:
key, value = filter_pair.split('=')
key = key.strip()
value = value.strip()
# Try to convert year to int
if key == "year":
try:
value = int(value)
except:
pass
filters[key] = value
return {
"semantic_query": semantic_query,
"filters": filters
}
# ============================================================================
# COMPONENT 4: ANSWER QUALITY CHECKER (FOR AGENTIC ROUTING)
# ============================================================================
@component
class AnswerQualityChecker:
"""Check if answer is satisfactory or needs web search fallback"""
@component.output_types(quality_score=str, route=str)
def run(self, answer: str, query: str) -> Dict[str, str]:
"""Check answer quality"""
# Simple heuristic - in production, use an LLM
if "I don't have" in answer or "cannot answer" in answer or len(answer) < 50:
return {"quality_score": "low", "route": "web_search"}
else:
return {"quality_score": "high", "route": "final_answer"}
# ============================================================================
# INDEXING WITH METADATA
# ============================================================================
def index_documents_with_metadata(document_store, file_path):
documents = load_documents_with_metadata(file_path)
embedder = SentenceTransformersDocumentEmbedder(model=EMBEDDING_MODEL)
embedder.warm_up()
docs_with_embeddings = embedder.run(documents)
document_store.write_documents(docs_with_embeddings["documents"])
print(f" Indexed {len(documents)} documents with metadata")
@component
class ReplySelector:
"""Select the primary reply from LLM output"""
@component.output_types(answer=str)
def run(self, replies: List[str]) -> Dict[str, str]:
if not replies:
return {"answer": ""}
return {"answer": replies[0]}
# ============================================================================
# BUILD EXPERT RAG PIPELINE WITH ALL TECHNIQUES
# ============================================================================
def build_expert_rag_pipeline(document_store):
"""Build expert RAG pipeline with HyDE, compression, and agentic routing"""
# Initialize LLMs
main_llm = OpenAIGenerator(
api_key=Secret.from_env_var("GROQ_API_KEY"),
api_base_url="https://api.groq.com/openai/v1",
model=GROQ_MODEL,
generation_kwargs={"max_tokens": 512, "temperature": 0.1}
)
hyde_llm = OpenAIGenerator(
api_key=Secret.from_env_var("GROQ_API_KEY"),
api_base_url="https://api.groq.com/openai/v1",
model=GROQ_MODEL,
generation_kwargs={"max_tokens": 300, "temperature": 0.1}
)
compressor_llm = OpenAIGenerator(
api_key=Secret.from_env_var("GROQ_API_KEY"),
api_base_url="https://api.groq.com/openai/v1",
model=GROQ_MODEL,
generation_kwargs={"max_tokens": 200, "temperature": 0.1}
)
# Initialize components
pipeline = Pipeline()
pipeline.add_component("reply_selector", ReplySelector())
# Self-querying
pipeline.add_component("self_query", SelfQueryAnalyzer(main_llm))
# HyDE generation
pipeline.add_component("hyde_generator", HyDEGenerator(hyde_llm))
# Embedders
pipeline.add_component(
"text_embedder",
SentenceTransformersTextEmbedder(model=EMBEDDING_MODEL)
)
pipeline.add_component(
"hyde_embedder",
SentenceTransformersTextEmbedder(model=EMBEDDING_MODEL)
)
# Dual retrievers
pipeline.add_component(
"bm25_retriever",
InMemoryBM25Retriever(document_store=document_store, top_k=BM25_TOP_K)
)
pipeline.add_component(
"semantic_retriever",
InMemoryEmbeddingRetriever(document_store=document_store, top_k=EMBEDDING_TOP_K)
)
# Document processing
pipeline.add_component("document_joiner", DocumentJoiner())
pipeline.add_component(
"ranker",
SentenceTransformersSimilarityRanker(model=RERANKER_MODEL, top_k=RERANKER_TOP_K)
)
# Contextual compression
pipeline.add_component("compressor", ContextualCompressor(compressor_llm))
# Answer generation
answer_template = """Answer the question based on the provided context.
If the context doesn't contain enough information, say so clearly.
Context:
{% for doc in documents %}
{{ doc.content }}
{% endfor %}
Question: {{ question }}
Answer:"""
pipeline.add_component("prompt_builder", PromptBuilder(template=answer_template))
pipeline.add_component("answer_generator", main_llm)
# Quality checker for agentic routing
pipeline.add_component("quality_checker", AnswerQualityChecker())
# Connect components
# Self-querying → retrieval
pipeline.connect("self_query.semantic_query", "bm25_retriever.query")
pipeline.connect("self_query.semantic_query", "text_embedder.text")
# HyDE pathway
pipeline.connect("self_query.semantic_query", "hyde_generator.query")
pipeline.connect("hyde_generator.hypothetical_doc", "hyde_embedder.text")
# Retrievers
pipeline.connect("text_embedder.embedding", "semantic_retriever.query_embedding")
# Join and rank
pipeline.connect("bm25_retriever.documents", "document_joiner.documents")
pipeline.connect("semantic_retriever.documents", "document_joiner.documents")
pipeline.connect("document_joiner.documents", "ranker.documents")
# Compression
pipeline.connect("ranker.documents", "compressor.documents")
# Answer generation
pipeline.connect("compressor.compressed_documents", "prompt_builder.documents")
pipeline.connect("prompt_builder", "answer_generator")
# Quality checking
pipeline.connect("answer_generator.replies", "reply_selector.replies")
pipeline.connect("reply_selector.answer", "quality_checker.answer")
return pipeline
# ============================================================================
# CONVERSATION CONTEXT MANAGER (FROM PART 2)
# ============================================================================
class ConversationContext:
"""Manages conversation history"""
def __init__(self, max_history=5):
self.history = []
self.max_history = max_history
def add_exchange(self, question: str, answer: str):
self.history.append({"question": question, "answer": answer})
if len(self.history) > self.max_history:
self.history = self.history[-self.max_history:]
def get_history_text(self) -> str:
if not self.history:
return ""
return "\n".join([f"Q: {e['question']}\nA: {e['answer']}" for e in self.history])
# ============================================================================
# EXPERT QUERY PROCESSOR
# ============================================================================
def process_expert_query(
pipeline,
query: str,
context: ConversationContext,
show_details=True
):
"""Process query with full expert RAG pipeline"""
if show_details:
print(f"\n{'='*70}")
print(f" Expert RAG Processing")
print(f"{'='*70}")
print(f"\n Original Query: {query}")
# Run the pipeline
try:
result = pipeline.run(
{
"self_query": {"query": query},
"ranker": {"query": query},
"compressor": {"query": query},
"prompt_builder": {"question": query},
"quality_checker": {"query": query}
},
include_outputs_from=["reply_selector", "quality_checker", "compressor"]
)
if show_details:
# Show self-querying results
if "self_query" in result:
print(f"\n Self-Query Analysis:")
print(f" Semantic Query: {result['self_query'].get('semantic_query', 'N/A')}")
print(f" Filters: {result['self_query'].get('filters', {})}")
# Show HyDE results
if "hyde_generator" in result:
hyde_doc = result['hyde_generator']['hypothetical_doc']
print(f"\n HyDE Hypothetical Document:")
print(f" {hyde_doc[:200]}...")
# Show compression results
if "compressor" in result:
print(f"\n Contextual Compression:")
compressed_docs = result['compressor']['compressed_documents']
print(f" Compressed {len(compressed_docs)} documents")
for i, doc in enumerate(compressed_docs[:2], 1):
print(f"\n Doc {i}: {doc.content[:150]}...")
# Show quality check
if "quality_checker" in result:
quality = result['quality_checker']
print(f"\n Answer Quality: {quality.get('quality_score', 'N/A')}")
print(f" Route Decision: {quality.get('route', 'N/A')}")
# Get final answer
answer = result["reply_selector"]["answer"]
if show_details:
print(f"\n{'─'*70}")
print(f" Final Answer:")
print(f"{answer}")
print(f"{'─'*70}")
# Update context
context.add_exchange(query, answer)
return answer, result
except Exception as e:
print(f"\n Error: {e}")
return f"Error processing query: {e}", {}
# ============================================================================
# MAIN EXECUTION
# ============================================================================
def main():
"""Main execution function"""
print("="*70)
print("EXPERT RAG SYSTEM - HAYSTACK + GROQ")
print(" Part 3: HyDE + Compression + Self-Querying + Agentic RAG")
print("="*70)
# Step 1: Initialize document store
print("\nStep 1: Initializing Document Store")
print("-"*70)
document_store = InMemoryDocumentStore()
# Step 2: Index documents with metadata
print("\nStep 2: Indexing Documents with Metadata")
print("-"*70)
index_documents_with_metadata(document_store, "/content/india_info.json")
# Step 3: Build expert RAG pipeline
print("\n Step 3: Building Expert RAG Pipeline")
print("-"*70)
pipeline = build_expert_rag_pipeline(document_store)
print("Expert RAG pipeline ready with:")
# Step 4: Initialize conversation context
print("\nStep 4: Initializing Conversation Context")
print("-"*70)
conversation = ConversationContext()
print("Conversation manager ready")
# Step 5: Test queries
print("\n" + "="*70)
print("TESTING EXPERT RAG")
print("="*70)
test_questions = [
"What are ISRO's major space achievements?",
"Tell me about India's economy",
"Find information about cricket from recent years",
"What makes the space program cost-effective?", # Follow-up with context
]
for i, question in enumerate(test_questions, 1):
print(f"\n\n{'#'*70}")
print(f"TEST QUERY {i}")
print(f"{'#'*70}")
answer, result = process_expert_query(
pipeline,
question,
conversation,
show_details=True
)
if __name__ == "__main__":
main()
实际应用场景
应用场景1：医疗问答系统
应用场景2：法律研究助手
应用场景3：客户支持机器人
完整的专家级RAG技术栈
Layer 1: Data Ingestion
- PDF/TXT/HTML converters
- Hierarchical chunking (better than fixed-size)
- Metadata extraction
Layer 2: Storage
- Vector DB (embeddings)
- Graph DB (relationships)
- SQL DB (metadata)
Layer 3: Retrieval
- HyDE generation
- Hybrid search (BM25 + Semantic)
- Self-querying with metadata
- Multi-hop retrieval
Layer 4: Processing
- Contextual compression
- Reranking
- Deduplication
Layer 5: Generation
- Agentic orchestration
- Tool usage
- Multi-source synthesis
- Source attribution
Layer 6: Monitoring
- Latency tracking
- Quality metrics
- Cost monitoring
- Error logging