Document Search

PageIndex enables reasoning-based RAG within a single document by default. For users who need to search across multiple documents, PageIndex provides three best-practice workflows for different scenarios.

Search Strategies

Choose the search strategy that best fits your use case:

Search by Metadata: For documents that can be distinguished by metadata (financial reports, legal documents, medical records)
Search by Semantics: For documents with different semantic content or covering diverse topics
Search by Description: A lightweight strategy for a small number of documents

Metadata Search
Semantic Search
Description Search

PageIndex with metadata support is in closed beta. Fill out this form to request early access to this feature.

When to Use

Use metadata search when your documents can be easily distinguished by structured attributes:

Financial reports categorized by company and time period
Legal documents categorized by case type
Medical records categorized by patient or condition

This method leverages “Query to SQL” for efficient document retrieval.

Implementation Pipeline

Generate PageIndex Trees

Upload all documents into PageIndex to get their doc_id.

# Upload documents and get doc_ids
doc_ids = []
for document in documents:
    response = pageindex.upload(document)
    doc_ids.append(response['doc_id'])

Set Up SQL Tables

Store documents along with their metadata and the PageIndex doc_id in a database table.

CREATE TABLE documents (
    id SERIAL PRIMARY KEY,
    doc_id VARCHAR(255) NOT NULL,
    doc_name VARCHAR(255),
    company VARCHAR(255),
    report_type VARCHAR(100),
    fiscal_year INTEGER,
    fiscal_quarter VARCHAR(10)
);

Query to SQL

Use an LLM to transform a user’s retrieval request into a SQL query to fetch relevant documents.

def query_to_sql(user_query: str, schema: str) -> str:
    prompt = f"""
    Given the database schema and user query, generate a SQL query.
    
    Schema:
    {schema}
    
    User Query: {user_query}
    
    Generate only the SQL query, no explanation.
    """
    return llm.generate(prompt)

# Example usage
sql_query = query_to_sql(
    "Find Apple's Q4 2023 financial reports",
    schema
)

Retrieve with PageIndex

Use the PageIndex doc_id of the retrieved documents to perform further retrieval via the PageIndex retrieval API.

# Execute SQL query to get relevant doc_ids
results = db.execute(sql_query)
doc_ids = [row['doc_id'] for row in results]

# Retrieve detailed information using PageIndex
for doc_id in doc_ids:
    response = pageindex.retrieve(
        doc_id=doc_id,
        query=user_query
    )
    print(response)

When to Use

Use semantic search for documents that cover diverse topics where vector-based similarity is effective. This method differs from classic vector search by incorporating document-level scoring.

Implementation Pipeline

Chunking and Embedding

Divide the documents into chunks, choose an embedding model to convert the chunks into vectors and store each vector with its corresponding doc_id in a vector database.

from openai import OpenAI
import chromadb

client = OpenAI()
chroma_client = chromadb.Client()
collection = chroma_client.create_collection("documents")

def chunk_document(text: str, chunk_size: int = 512) -> list[str]:
    """Split document into chunks"""
    words = text.split()
    chunks = []
    for i in range(0, len(words), chunk_size):
        chunk = ' '.join(words[i:i + chunk_size])
        chunks.append(chunk)
    return chunks

# Process documents
for doc_id, document in documents.items():
    chunks = chunk_document(document)
    
    # Generate embeddings
    embeddings = client.embeddings.create(
        input=chunks,
        model="text-embedding-3-small"
    )
    
    # Store in vector database
    for i, (chunk, embedding) in enumerate(zip(chunks, embeddings.data)):
        collection.add(
            ids=[f"{doc_id}_chunk_{i}"],
            embeddings=[embedding.embedding],
            metadatas=[{"doc_id": doc_id, "chunk_index": i}],
            documents=[chunk]
        )

Vector Search

For each query, conduct a vector-based search to get top-K chunks with their corresponding documents.

def vector_search(query: str, top_k: int = 20):
    # Generate query embedding
    query_embedding = client.embeddings.create(
        input=[query],
        model="text-embedding-3-small"
    )
    
    # Search vector database
    results = collection.query(
        query_embeddings=[query_embedding.data[0].embedding],
        n_results=top_k
    )
    
    return results

Compute Document Score

For each document, calculate a relevance score using the following formula:

\text{DocScore}=\frac{1}{\sqrt{N+1}}\sum_{n=1}^N \text{ChunkScore}(n)

Where:

N is the number of content chunks associated with each document
ChunkScore(n) is the relevance score of chunk n
The sum aggregates relevance from all related chunks
The +1 inside the square root ensures the formula handles nodes with zero chunks
Using the square root in the denominator allows the score to increase with the number of relevant chunks, but with diminishing returns

This scoring favors documents with fewer, highly relevant chunks over those with many weakly relevant ones.

import math
from collections import defaultdict

def compute_document_scores(search_results):
    # Group chunks by document
    doc_chunks = defaultdict(list)
    for i, metadata in enumerate(search_results['metadatas'][0]):
        doc_id = metadata['doc_id']
        score = search_results['distances'][0][i]
        doc_chunks[doc_id].append(score)
    
    # Calculate document scores
    doc_scores = {}
    for doc_id, chunk_scores in doc_chunks.items():
        N = len(chunk_scores)
        score_sum = sum(chunk_scores)
        doc_scores[doc_id] = score_sum / math.sqrt(N + 1)
    
    return sorted(doc_scores.items(), key=lambda x: x[1], reverse=True)

Retrieve with PageIndex

Select the documents with the highest DocScore, then use their doc_id to perform further retrieval via the PageIndex retrieval API.

# Get vector search results
search_results = vector_search(user_query)

# Compute document scores
ranked_docs = compute_document_scores(search_results)

# Retrieve top documents using PageIndex
top_n = 3
for doc_id, score in ranked_docs[:top_n]:
    response = pageindex.retrieve(
        doc_id=doc_id,
        query=user_query
    )
    print(f"Document: {doc_id} (Score: {score})")
    print(response)

When to Use

Use description-based search for documents that don’t have metadata. This is a lightweight approach that works best with a small number of documents.

Implementation Pipeline

Generate PageIndex Trees

Upload all documents into PageIndex to get their doc_id and tree structure.

# Upload documents and retrieve tree structures
document_trees = {}
for document in documents:
    response = pageindex.upload(document)
    doc_id = response['doc_id']
    
    # Get tree structure
    tree = pageindex.get_tree(doc_id)
    document_trees[doc_id] = {
        'name': document.name,
        'tree': tree
    }

Generate Descriptions

Generate a description for each document based on its PageIndex tree structure and node summaries.

def generate_description(tree_structure: dict) -> str:
    prompt = f"""
    You are given a table of contents structure of a document. 
    Your task is to generate a one-sentence description for the document that makes it easy to distinguish from other documents.
        
    Document tree structure: {tree_structure}
    
    Directly return the description, do not include any other text.
    """
    
    return llm.generate(prompt)

# Generate descriptions for all documents
descriptions = {}
for doc_id, doc_info in document_trees.items():
    description = generate_description(doc_info['tree'])
    descriptions[doc_id] = {
        'name': doc_info['name'],
        'description': description
    }

Search with LLM

Use an LLM to select relevant documents by comparing the user query against the generated descriptions.

def search_by_description(query: str, descriptions: dict) -> list[str]:
    # Prepare documents list
    docs_list = [
        {
            "doc_id": doc_id,
            "doc_name": info['name'],
            "doc_description": info['description']
        }
        for doc_id, info in descriptions.items()
    ]
    
    prompt = f""" 
    You are given a list of documents with their IDs, file names, and descriptions. Your task is to select documents that may contain information relevant to answering the user query.
    
    Query: {query}
    
    Documents: {docs_list}
    
    Response Format:
    {{
        "thinking": "<Your reasoning for document selection>",
        "answer": <Python list of relevant doc_ids>, e.g. ['doc_id1', 'doc_id2']. Return [] if no documents are relevant.
    }}
    
    Return only the JSON structure, with no additional output.
    """
    
    response = llm.generate(prompt)
    result = json.loads(response)
    return result['answer']

Retrieve with PageIndex

Use the PageIndex doc_id of the retrieved documents to perform further retrieval via the PageIndex retrieval API.

# Search for relevant documents
relevant_doc_ids = search_by_description(user_query, descriptions)

# Retrieve detailed information using PageIndex
for doc_id in relevant_doc_ids:
    response = pageindex.retrieve(
        doc_id=doc_id,
        query=user_query
    )
    print(f"Document: {descriptions[doc_id]['name']}")
    print(response)

Get Started

Core Concepts

Guides

Cookbook

Tutorials

Search Strategies

When to Use

Implementation Pipeline

When to Use

Implementation Pipeline

When to Use

Implementation Pipeline

Need Help?

Build docs developers (and LLMs) love

Get Started

Core Concepts

Guides

Cookbook

Tutorials

​Search Strategies

​When to Use

​Implementation Pipeline

​When to Use

​Implementation Pipeline

​When to Use

​Implementation Pipeline

​Need Help?

Build docs developers (and LLMs) love

Search Strategies

When to Use

Implementation Pipeline

When to Use

Implementation Pipeline

When to Use

Implementation Pipeline

Need Help?