import os
from typing import List, Dict, Tuple, Optional
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.document_loaders import PyPDFLoader
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import Chroma
from langchain_community.chat_models import ChatOpenAI
from langchain.chains import create_extraction_chain
from langchain.prompts import PromptTemplate
from dataclasses import dataclass
import uuid
import json
from anthropic import Anthropic
import numpy as np
from rank_bm25 import BM25Okapi
import logging
from cohere import Client
def reciprocal_rank_fusion(result_lists, weights=None):
"""Combine multiple ranked lists using reciprocal rank fusion"""
fused_scores = {}
num_lists = len(result_lists)
if weights is None:
weights = [1.0] * num_lists
for i in range(num_lists):
for doc_id, score in result_lists[i]:
if doc_id not in fused_scores:
fused_scores[doc_id] = 0
fused_scores[doc_id] += weights[i] * score
# Sort by score in descending order
sorted_results = sorted(
fused_scores.items(),
key=lambda x: x[1],
reverse=True
)
return sorted_results
os.environ["LANGCHAIN_TRACING_V2"]="true"
os.environ["LANGCHAIN_ENDPOINT"]="https://api.smith.langchain.com"
os.environ.get("LANGCHAIN_API_KEY")
os.environ["LANGCHAIN_PROJECT"]="VELLA"
@dataclass
class DocumentChunk:
content: str
page_number: int
chunk_id: str
start_char: int
end_char: int
@dataclass
class RetrievalConfig:
num_chunks: int = 5
embedding_weight: float = 0.5
bm25_weight: float = 0.5
context_window: int = 3
chunk_overlap: int = 200
chunk_size: int = 1000
@dataclass
class ContextualizedChunk(DocumentChunk):
context: str = ""
embedding: Optional[np.ndarray] = None
bm25_score: Optional[float] = None
class DocumentSummarizer:
def __init__(self, openai_api_key: str, cohere_api_key: str, embedding_model, chunk_size, chunk_overlap, num_k_rerank, model_cohere_rerank):
self.openai_api_key = openai_api_key
self.cohere_client = Client(cohere_api_key)
self.embeddings = HuggingFaceEmbeddings(
model_name=embedding_model
)
self.text_splitter = RecursiveCharacterTextSplitter(
chunk_size=chunk_size,
chunk_overlap=chunk_overlap
)
self.chunk_metadata = {} # Store chunk metadata for tracing
self.num_k_rerank = num_k_rerank
self.model_cohere_rerank = model_cohere_rerank
def load_and_split_document(self, pdf_path: str) -> List[DocumentChunk]:
"""Load PDF and split into chunks with metadata"""
loader = PyPDFLoader(pdf_path)
pages = loader.load()
chunks = []
char_count = 0
for page in pages:
text = page.page_content
# Split the page content
page_chunks = self.text_splitter.split_text(text)
for chunk in page_chunks:
chunk_id = str(uuid.uuid4())
start_char = text.find(chunk)
end_char = start_char + len(chunk)
doc_chunk = DocumentChunk(
content=chunk,
page_number=page.metadata.get('page') + 1, # 1-based page numbering
chunk_id=chunk_id,
start_char=char_count + start_char,
end_char=char_count + end_char
)
chunks.append(doc_chunk)
# Store metadata for later retrieval
self.chunk_metadata[chunk_id] = {
'page': doc_chunk.page_number,
'start_char': doc_chunk.start_char,
'end_char': doc_chunk.end_char
}
char_count += len(text)
return chunks
def create_vector_store(self, chunks: List[DocumentChunk]) -> Chroma:
"""Create vector store with metadata"""
texts = [chunk.content for chunk in chunks]
metadatas = [{
'chunk_id': chunk.chunk_id,
'page': chunk.page_number,
'start_char': chunk.start_char,
'end_char': chunk.end_char
} for chunk in chunks]
vector_store = Chroma.from_texts(
texts=texts,
metadatas=metadatas,
embedding=self.embeddings
)
return vector_store
def rerank_chunks(
self,
chunks: List[Dict],
query: str,
k: int = 5
) -> List[Dict]:
"""
Rerank chunks using Cohere's reranking model.
Args:
chunks: List of dictionaries containing chunks and their metadata
query: Original search query
k: Number of top chunks to return
Returns:
List of reranked chunks with updated relevance scores
"""
try:
# Prepare documents for reranking
documents = [chunk['content'] for chunk in chunks]
# Get reranking scores from Cohere
results = self.cohere_client.rerank(
query=query,
documents=documents,
top_n=k,
model=self.model_cohere_rerank
)
# Create reranked results with original metadata
reranked_chunks = []
for hit in results:
original_chunk = chunks[hit.index]
reranked_chunks.append({
**original_chunk,
'relevance_score': hit.relevance_score
})
return reranked_chunks
except Exception as e:
logging.error(f"Reranking failed: {str(e)}")
return chunks[:k] # Fallback to original ordering
def generate_summary_with_sources(
self,
vector_store: Chroma,
query: str = "Summarize the main points of this document"
) -> List[Dict]:
"""Generate summary with source citations using reranking"""
# Retrieve more initial chunks for reranking
relevant_docs = vector_store.similarity_search_with_score(query, k=20)
# Prepare chunks for reranking
chunks = []
for doc, score in relevant_docs:
chunks.append({
'content': doc.page_content,
'page': doc.metadata['page'],
'chunk_id': doc.metadata['chunk_id'],
'relevance_score': score
})
# Rerank chunks
reranked_chunks = self.rerank_chunks(chunks, query, k=self.num_k_rerank)
# Prepare context and sources from reranked chunks
contexts = []
sources = []
for chunk in reranked_chunks:
contexts.append(chunk['content'])
sources.append({
'content': chunk['content'],
'page': chunk['page'],
'chunk_id': chunk['chunk_id'],
'relevance_score': chunk['relevance_score']
})
prompt_template = """
Based on the following context, provide multiple key points from the document.
For each point, create a new paragraph.
Each paragraph should be a complete, self-contained insight.
Context: {context}
Key points:
"""
prompt = PromptTemplate(
template=prompt_template,
input_variables=["context"]
)
llm = ChatOpenAI(
temperature=0,
model_name="gpt-4o-mini",
api_key=self.openai_api_key
)
response = llm.predict(prompt.format(context="\n\n".join(contexts)))
# Split the response into paragraphs
summaries = [p.strip() for p in response.split('\n\n') if p.strip()]
# Create structured output
structured_output = []
for idx, summary in enumerate(summaries):
# Associate each summary with the most relevant source
structured_output.append({
"content": summary,
"source": {
"page": sources[min(idx, len(sources)-1)]['page'],
"text": sources[min(idx, len(sources)-1)]['content'][:200] + "...",
"relevance_score": sources[min(idx, len(sources)-1)]['relevance_score']
}
})
return structured_output
def get_source_context(self, chunk_id: str, window: int = 100) -> Dict:
"""Get extended context around a specific chunk"""
metadata = self.chunk_metadata.get(chunk_id)
if not metadata:
return None
return {
'page': metadata['page'],
'start_char': metadata['start_char'],
'end_char': metadata['end_char']
}
class ContextualRetriever:
def __init__(self, config: RetrievalConfig, claude_api_key: str, claude_context_model):
self.config = config
self.claude_client = Anthropic(api_key=claude_api_key)
self.logger = logging.getLogger(__name__)
self.bm25 = None
self.claude_context_model = claude_context_model
def generate_context(self, full_text: str, chunk: DocumentChunk) -> str:
"""Generate contextual description using Claude"""
try:
prompt = f"""
{full_text}
Here is the chunk we want to situate within the whole document
{chunk.content}
Please give a short succinct context to situate this chunk within the overall document for the purposes of improving search retrieval of the chunk. Answer only with the succinct context and nothing else."""
response = self.claude_client.messages.create(
model=self.claude_context_model,
max_tokens=100,
messages=[{"role": "user", "content": prompt}]
)
return response.content[0].text
except Exception as e:
self.logger.error(f"Context generation failed for chunk {chunk.chunk_id}: {str(e)}")
return ""
def contextualize_chunks(self, full_text: str, chunks: List[DocumentChunk]) -> List[ContextualizedChunk]:
"""Add context to all chunks"""
contextualized_chunks = []
for chunk in chunks:
context = self.generate_context(full_text, chunk)
contextualized_chunk = ContextualizedChunk(
content=chunk.content,
page_number=chunk.page_number,
chunk_id=chunk.chunk_id,
start_char=chunk.start_char,
end_char=chunk.end_char,
context=context
)
contextualized_chunks.append(contextualized_chunk)
return contextualized_chunks
class EnhancedDocumentSummarizer(DocumentSummarizer):
def __init__(self, openai_api_key: str, claude_api_key: str, config: RetrievalConfig, embedding_model, chunk_size, chunk_overlap, num_k_rerank, model_cohere_rerank, claude_context_model, system_prompt, gpt_model, gpt_temperature):
super().__init__(openai_api_key, os.environ.get("COHERE_API_KEY"), embedding_model, chunk_size, chunk_overlap, num_k_rerank, model_cohere_rerank)
self.config = config
self.contextual_retriever = ContextualRetriever(config, claude_api_key, claude_context_model)
self.logger = logging.getLogger(__name__)
self.system_prompt = system_prompt
self.gpt_model = gpt_model
self.gpt_temperature = gpt_temperature
def create_enhanced_vector_store(self, chunks: List[ContextualizedChunk]) -> Tuple[Chroma, BM25Okapi, List[str]]:
"""Create vector store and BM25 index with contextualized chunks"""
try:
# Prepare texts with context
texts = [f"{chunk.context} {chunk.content}" for chunk in chunks]
# Create vector store
metadatas = [{
'chunk_id': chunk.chunk_id,
'page': chunk.page_number,
'start_char': chunk.start_char,
'end_char': chunk.end_char,
'context': chunk.context
} for chunk in chunks]
vector_store = Chroma.from_texts(
texts=texts,
metadatas=metadatas,
embedding=self.embeddings
)
# Create BM25 index
tokenized_texts = [text.split() for text in texts]
bm25 = BM25Okapi(tokenized_texts)
# Get chunk IDs in order
chunk_ids = [chunk.chunk_id for chunk in chunks]
return vector_store, bm25, chunk_ids
except Exception as e:
self.logger.error(f"Error creating enhanced vector store: {str(e)}")
raise
def retrieve_with_rank_fusion(
self,
vector_store: Chroma,
bm25: BM25Okapi,
chunk_ids: List[str],
query: str
) -> List[Dict]:
"""Combine embedding and BM25 retrieval results"""
try:
# Get embedding results
embedding_results = vector_store.similarity_search_with_score(
query,
k=self.config.num_chunks
)
# Convert embedding results to list of (chunk_id, score)
embedding_list = [
(doc.metadata['chunk_id'], 1 / (1 + score))
for doc, score in embedding_results
]
# Get BM25 results
tokenized_query = query.split()
bm25_scores = bm25.get_scores(tokenized_query)
# Convert BM25 scores to list of (chunk_id, score)
bm25_list = [
(chunk_ids[i], float(score)) for i, score in enumerate(bm25_scores)
]
# Sort bm25_list by score in descending order and limit to top N results
bm25_list = sorted(bm25_list, key=lambda x: x[1], reverse=True)[:self.config.num_chunks]
# Normalize BM25 scores
max_bm25 = max([score for _, score in bm25_list]) if bm25_list else 1
bm25_list = [(doc_id, score / max_bm25) for doc_id, score in bm25_list]
# Pass the lists to rank fusion
result_lists = [embedding_list, bm25_list]
weights = [self.config.embedding_weight, self.config.bm25_weight]
combined_results = reciprocal_rank_fusion(
result_lists,
weights=weights
)
return combined_results
except Exception as e:
self.logger.error(f"Error in rank fusion retrieval: {str(e)}")
raise
def generate_enhanced_summary(
self,
vector_store: Chroma,
bm25: BM25Okapi,
chunk_ids: List[str],
query: str = "Summarize the main points of this document"
) -> List[Dict]:
"""Generate enhanced summary using both vector and BM25 retrieval"""
try:
# Get combined results using rank fusion
ranked_results = self.retrieve_with_rank_fusion(
vector_store,
bm25,
chunk_ids,
query
)
# Prepare context and track sources
contexts = []
sources = []
# Get full documents for top results
for chunk_id, score in ranked_results[:self.config.num_chunks]:
results = vector_store.get(
where={"chunk_id": chunk_id},
include=["documents", "metadatas"]
)
if results["documents"]:
context = results["documents"][0]
metadata = results["metadatas"][0]
contexts.append(context)
sources.append({
'content': context,
'page': metadata['page'],
'chunk_id': chunk_id,
'relevance_score': score,
'context': metadata.get('context', '')
})
prompt_template = self.system_prompt
prompt = PromptTemplate(
template=prompt_template,
input_variables=["context"]
)
llm = ChatOpenAI(
temperature=self.gpt_temperature,
model_name=self.gpt_model,
api_key=self.openai_api_key,
)
response = llm.predict(prompt.format(context="\n\n".join(contexts)))
# Split the response into paragraphs
summaries = [p.strip() for p in response.split('\n\n') if p.strip()]
# Create structured output
structured_output = []
for idx, summary in enumerate(summaries):
source_idx = min(idx, len(sources)-1)
structured_output.append({
"content": summary,
"source": {
"page": sources[source_idx]['page'],
"text": sources[source_idx]['content'][:200] + "...",
"context": sources[source_idx]['context'],
"relevance_score": sources[source_idx]['relevance_score'],
"chunk_id": sources[source_idx]['chunk_id']
}
})
return structured_output
except Exception as e:
self.logger.error(f"Error generating enhanced summary: {str(e)}")
raise
def get_llm_summary_answer_by_cursor_complete(serializer, listaPDFs):
allPdfsChunks = []
# Configuration
config = RetrievalConfig(
num_chunks=serializer["num_chunks_retrieval"],
embedding_weight=serializer["embedding_weight"],
bm25_weight=serializer["bm25_weight"],
context_window=serializer["context_window"],
chunk_overlap=serializer["chunk_overlap"]
)
# Initialize enhanced summarizer
summarizer = EnhancedDocumentSummarizer(
openai_api_key=os.environ.get("OPENAI_API_KEY"),
claude_api_key= os.environ.get("CLAUDE_API_KEY"),
config=config,
embedding_model=serializer["hf_embedding"],
chunk_overlap=serializer["chunk_overlap"],
chunk_size=serializer["chunk_size"],
num_k_rerank=serializer["num_k_rerank"],
model_cohere_rerank=serializer["model_cohere_rerank"],
claude_context_model=serializer["claude_context_model"],
system_prompt=serializer["system_prompt"],
gpt_model=serializer["model"],
gpt_temperature=serializer["gpt_temperature"]
)
# # Load and process document
# pdf_path = "./Im_a_storyteller.pdf"
# chunks = summarizer.load_and_split_document(pdf_path)
# Load and process document
for pdf in listaPDFs:
pdf_path = pdf
chunks = summarizer.load_and_split_document(pdf_path)
allPdfsChunks = allPdfsChunks + chunks
# Get full text for contextualization
loader = PyPDFLoader(pdf_path)
pages = loader.load()
full_text = " ".join([page.page_content for page in pages])
# Contextualize chunks
contextualized_chunks = summarizer.contextual_retriever.contextualize_chunks(full_text, allPdfsChunks)
# Create enhanced vector store and BM25 index
vector_store, bm25, chunk_ids = summarizer.create_enhanced_vector_store(contextualized_chunks)
# Generate enhanced summary
structured_summaries = summarizer.generate_enhanced_summary(
vector_store,
bm25,
chunk_ids,
serializer["user_message"]
)
# Output results as JSON
json_output = json.dumps(structured_summaries, indent=2)
print("\nStructured Summaries:")
print(json_output)
return {
"resultado": structured_summaries,
"parametros-utilizados": {
"num_chunks_retrieval": serializer["num_chunks_retrieval"],
"embedding_weight": serializer["embedding_weight"],
"bm25_weight": serializer["bm25_weight"],
"context_window": serializer["context_window"],
"chunk_overlap": serializer["chunk_overlap"],
"num_k_rerank": serializer["num_k_rerank"],
"model_cohere_rerank": serializer["model_cohere_rerank"],
"more_initial_chunks_for_reranking": serializer["more_initial_chunks_for_reranking"],
"claude_context_model": serializer["claude_context_model"],
"gpt_temperature": serializer["gpt_temperature"],
"user_message": serializer["user_message"],
"model": serializer["model"],
"hf_embedding": serializer["hf_embedding"],
"chunk_size": serializer["chunk_size"],
"chunk_overlap": serializer["chunk_overlap"],
"system_prompt": serializer["system_prompt"],
}}