Anthropic提出的Contextual RAG开源实现Open Contextual RAG来了

!pip install together # To access open source LLMs
!pip install --upgrade tiktoken # To count total token counts
!pip install beautifulsoup4 # To scrape documents to RAG over
!pip install bm25s # To implement out key-word BM25 search
1.
2.
3.
4.

# Let's download the essay from Paul Graham's website

import requests
from bs4 import BeautifulSoup

def scrape_pg_essay():

    url = 'https://paulgraham.com/foundermode.html'

    try:
        # Send GET request to the URL
        response = requests.get(url)
        response.raise_for_status()  # Raise an error for bad status codes

        # Parse the HTML content
        soup = BeautifulSoup(response.text, 'html.parser')

        # Paul Graham's essays typically have the main content in a font tag
        # You might need to adjust this selector based on the actual HTML structure
        content = soup.find('font')

        if content:
            # Extract and clean the text
            text = content.get_text()
            # Remove extra whitespace and normalize line breaks
            text = ' '.join(text.split())
            return text
        else:
            return "Could not find the main content of the essay."

    except requests.RequestException as e:
        return f"Error fetching the webpage: {e}"

# Scrape the essay
pg_essay = scrape_pg_essay()

def create_chunks(document, chunk_size=300, overlap=50):
    return [document[i : i + chunk_size] for i in range(0, len(document), chunk_size - overlap)]
     

chunks = create_chunks(pg_essay, chunk_size=250, overlap=30)

for i, chunk in enumerate(chunks):
    print(f"Chunk {i + 1}: {chunk}")
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.

import together, os
from together import Together

# Paste in your Together AI API Key or load it
TOGETHER_API_KEY = os.environ.get("TOGETHER_API_KEY")
     

# We want to generate a snippet explaining the relevance/importance of the chunk with
# full document in mind.

CONTEXTUAL_RAG_PROMPT = """
Given the document below, we want to explain what the chunk captures in the document.

{WHOLE_DOCUMENT}

Here is the chunk we want to explain:

{CHUNK_CONTENT}

Answer ONLY with a succinct explaination of the meaning of the chunk in the context of the whole document above.
"""
     

from typing import List

# First we will just generate the prompts and examine them

def generate_prompts(document : str, chunks : List[str]) -> List[str]:
  prompts = []
  for chunk in chunks:
    prompt = CONTEXTUAL_RAG_PROMPT.format(WHOLE_DOCUMENT=document, CHUNK_CONTENT=chunk)
    prompts.append(prompt)
  return prompts
     

prompts = generate_prompts(pg_essay, chunks)
     
def generate_context(prompt: str):
    """
    Generates a contextual response based on the given prompt using the specified language model.
    Args:
        prompt (str): The input prompt to generate a response for.
    Returns:
        str: The generated response content from the language model.
    """
    response = client.chat.completions.create(
        model="meta-llama/Llama-3.2-3B-Instruct-Turbo",
        messages=[{"role": "user", "content": prompt}],
        temperature=1
    )
    return response.choices[0].message.content
     
# Let's generate the entire list of contextual chunks

contextual_chunks = [generate_context(prompts[i])+' '+chunks[i] for i in range(len(chunks))]
     

# We should have one contextual chunk for each original chunk
len(contextual_chunks), len(contextual_chunks) == len(chunks)
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.

from typing import List
import together
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

def generate_embeddings(input_texts: List[str], model_api_string: str) -> List[List[float]]:
    """Generate embeddings from Together python library.

    Args:
        input_texts: a list of string input texts.
        model_api_string: str. An API string for a specific embedding model of your choice.

    Returns:
        embeddings_list: a list of embeddings. Each element corresponds to the each input text.
    """
    outputs = client.embeddings.create(
        input=input_texts,
        model=model_api_string,
    )
    return [x.embedding for x in outputs.data]
     

contextual_embeddings = generate_embeddings(contextual_chunks, "BAAI/bge-large-en-v1.5")
     

def vector_retreival(query: str, top_k: int = 5, vector_index: np.ndarray = None) -> List[int]:
    """
    Retrieve the top-k most similar items from an index based on a query.
    Args:
        query (str): The query string to search for.
        top_k (int, optional): The number of top similar items to retrieve. Defaults to 5.
        index (np.ndarray, optional): The index array containing embeddings to search against. Defaults to None.
    Returns:
        List[int]: A list of indices corresponding to the top-k most similar items in the index.
    """

    query_embedding = generate_embeddings([query], 'BAAI/bge-large-en-v1.5')[0]
    similarity_scores = cosine_similarity([query_embedding], vector_index)

    return list(np.argsort(-similarity_scores)[0][:top_k])
     

vector_retreival(query = "What are 'skip-level' meetings?", top_k = 5, vector_index = contextual_embeddings)
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.

import bm25s

# Create the BM25 model and index the corpus
retriever = bm25s.BM25(corpus=contextual_chunks)
retriever.index(bm25s.tokenize(contextual_chunks))

# Query the corpus and get top-k results
query = "What are 'skip-level' meetings?"
results, scores = retriever.retrieve(bm25s.tokenize(query), k=5,)
     
for doc in results[0]:
  print(f"Chunk Index {contextual_chunks.index(doc)} : {doc}")

def bm25_retreival(query: str, k : int, bm25_index) -> List[int]:
    """
    Retrieve the top-k document indices based on the BM25 algorithm for a given query.
    Args:
        query (str): The search query string.
        k (int): The number of top documents to retrieve.
        bm25_index: The BM25 index object used for retrieval.
    Returns:
        List[int]: A list of indices of the top-k documents that match the query.
    """

    results, scores = bm25_index.retrieve(bm25s.tokenize(query), k=k)

    return [contextual_chunks.index(doc) for doc in results[0]]
     

bm25_retreival(query = "What are 'skip-level' meetings?", k = 5, bm25_index = retriever)
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.

from collections import defaultdict

def reciprocal_rank_fusion(*list_of_list_ranks_system, K=60):
    """
    Fuse rank from multiple IR systems using Reciprocal Rank Fusion.

    Args:
    * list_of_list_ranks_system: Ranked results from different IR system.
    K (int): A constant used in the RRF formula (default is 60).

    Returns:
    Tuple of list of sorted documents by score and sorted documents
    """
    # Dictionary to store RRF mapping
    rrf_map = defaultdict(float)

    # Calculate RRF score for each result in each list
    for rank_list in list_of_list_ranks_system:
        for rank, item in enumerate(rank_list, 1):
            rrf_map[item] += 1 / (rank + K)

    # Sort items based on their RRF scores in descending order
    sorted_items = sorted(rrf_map.items(), key=lambda x: x[1], reverse=True)

    # Return tuple of list of sorted documents by score and sorted documents
    return sorted_items, [item for item, score in sorted_items]

# Example ranked lists from different sources
vector_top_k = vector_retreival(query = "What are 'skip-level' meetings?", top_k = 5, vector_index = contextual_embeddings)
bm25_top_k = bm25_retreival(query = "What are 'skip-level' meetings?", k = 5, bm25_index = retriever)

# Combine the lists using RRF
hybrid_top_k = reciprocal_rank_fusion(vector_top_k, bm25_top_k)

for index in hybrid_top_k[1]:
  print(f"Chunk Index {index} : {contextual_chunks[index]}")
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.

from together import Together

query = "What are 'skip-level' meetings?" # we keep the same query - can change if we want

response = client.rerank.create(
  model="Salesforce/Llama-Rank-V1",
  query=query,
  documents=hybrid_top_k_docs,
  top_n=3 # we only want the top 3 results
)

for result in response.results:
    print(f"Document: {hybrid_top_k_docs[result.index]}")
    print(f"Relevance Score: {result.relevance_score}\n\n")
    
# Document: This chunk refers to "skip-level" meetings, which are a key characteristic of founder mode, where the CEO engages directly with the company beyond their direct reports. This contrasts with the "manager mode" of addressing company issues, where decisions are made perfunctorily via a hierarchical system, to which founders instinctively rebel. that there's a name for it. And once you abandon that constraint there are a huge number of permutations to choose from.For example, Steve Jobs used to run an annual retreat for what he considered the 100 most important people at Apple, and these wer
# Relevance Score: 0.9733742140554578

# Document: This chunk discusses the shift in company management away from the "manager mode" that most companies follow, where CEOs engage with the company only through their direct reports, to "founder mode", where CEOs engage more directly with even higher-level employees and potentially skip over direct reports, potentially leading to "skip-level" meetings. ts of, it's pretty clear that it's going to break the principle that the CEO should engage with the company only via his or her direct reports. "Skip-level" meetings will become the norm instead of a practice so unusual that there's a name for it. An
# Relevance Score: 0.7481956787777269

# Document: This chunk explains that founder mode, a hypothetical approach to running a company by its founders, will differ from manager mode in that founders will engage directly with the company, rather than just their direct reports, through "skip-level" meetings, disregarding the traditional principle that CEOs should only interact with their direct reports, as managers do.  can already guess at some of the ways it will differ.The way managers are taught to run companies seems to be like modular design in the sense that you treat subtrees of the org chart as black boxes. You tell your direct reports what to do, and it's
# Relevance Score: 0.7035533028052896

# Lets add the top 3 documents to a string

retreived_chunks = ''

for result in response.results:
    retreived_chunks += hybrid_top_k_docs[result.index] + '\n\n'

print(retreived_chunks)

# This chunk refers to "skip-level" meetings, which are a key characteristic of founder mode, where the CEO engages directly with the company beyond their direct reports. This contrasts with the "manager mode" of addressing company issues, where decisions are made perfunctorily via a hierarchical system, to which founders instinctively rebel. that there's a name for it. And once you abandon that constraint there are a huge number of permutations to choose from.For example, Steve Jobs used to run an annual retreat for what he considered the 100 most important people at Apple, and these wer

# This chunk discusses the shift in company management away from the "manager mode" that most companies follow, where CEOs engage with the company only through their direct reports, to "founder mode", where CEOs engage more directly with even higher-level employees and potentially skip over direct reports, potentially leading to "skip-level" meetings. ts of, it's pretty clear that it's going to break the principle that the CEO should engage with the company only via his or her direct reports. "Skip-level" meetings will become the norm instead of a practice so unusual that there's a name for it. An

# This chunk explains that founder mode, a hypothetical approach to running a company by its founders, will differ from manager mode in that founders will engage directly with the company, rather than just their direct reports, through "skip-level" meetings, disregarding the traditional principle that CEOs should only interact with their direct reports, as managers do.  can already guess at some of the ways it will differ.The way managers are taught to run companies seems to be like modular design in the sense that you treat subtrees of the org chart as black boxes. You tell your direct reports what to do, and it's
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.

# Generate a story based on the top 10 most similar movies

query = "What are 'skip-level' meetings?"

response = client.chat.completions.create(
    model="meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo",
    messages=[
      {"role": "system", "content": "You are a helpful chatbot."},
      {"role": "user", "content": f"Answer the question: {query}. Here is relevant information: {retreived_chunks}"},
    ],
)
     

response.choices[0].message.content
     
# '"Skip-level" meetings refer to a management practice where a CEO or high-level executive engages directly with employees who are not their direct reports, bypassing the traditional hierarchical structure of the organization. This approach is characteristic of "founder mode," where the CEO seeks to have a more direct connection with the company beyond their immediate team. In contrast to the traditional "manager mode," where decisions are made through a hierarchical system, skip-level meetings allow for more open communication and collaboration between the CEO and various levels of employees. This approach is often used by founders who want to stay connected to the company\'s operations and culture, and to foster a more flat and collaborative organizational structure.'
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.