Vectorless, Reasoning-based RAG
for long, complex documents
Retrieval that's accurate, traceable, and context-aware.
No vector DB, no embeddings, no chunking.
<cite doc="report.pdf" page="43"/>Get Started in Seconds
Install the SDK
pip install -U pageindexGet a PageIndex API key
Create one in the Developer Dashboard.
Index a document
Submit a PDF, and PageIndex builds a tree index that preserves the document's structure.
Ask a question
Your LLM agentically reasons over the tree to retrieve relevant information and form answers.
import osfrom pageindex import PageIndexClient os.environ["PAGEINDEX_API_KEY"] = "your-pageindex-key"os.environ["OPENAI_API_KEY"] = "your-openai-api-key" client = PageIndexClient( index="cloud", # index and store in PageIndex Cloud chat="gpt-5.6-sol", # your own LLM answers) doc_id = client.submit_document("report.pdf", wait=True)["doc_id"] query = "What benchmarks are used?" for chunk in client.chat(query, doc_id=doc_id, stream=True): print(chunk, end="", flush=True) Proven on Benchmarks
Try it on your documentsFinanceBench is the industry-standard QA benchmark over SEC filings, and the questions need the right pages out of long, dense reports.
On it, PageIndex reaches 98.7% retrieval accuracy — still the highest to date.
Run It Local or on Cloud
PageIndex Local
Open source, on your own machine, with your own model key.
client = PageIndexClient( index="gpt-5.6-luna", # PageIndex Flash, running locally)- Documents and tree indexes never leave your disk
- PageIndex Flash indexes a 1,000-page PDF in minutes, for about a dollar
- Built for text-based PDFs — reads the layout directly, no OCR
- Fully open source, 35k+ stars on GitHub
PageIndex Cloud
Fully managed indexing, storage, and retrieval. Nothing to run.
client = PageIndexClient( index="cloud", # index and store in PageIndex Cloud)- OCR and image understanding for scanned and image-heavy files
- Hosted document storage and retrieval API, scaling handled for you
- API keys, usage, and indexed documents in one dashboard
- Free tier to start — 30k+ developers already building on it
PageIndex vs Vector DB
PageIndex
Logical Reasoning
High Retrieval Accuracy
Relies on logical reasoning to determine relevance rather than similarity, ideal for domain-specific data.
Context-Aware Retrieval
Retrieval depends on full context (e.g., conversational history and domain or enterprise knowledge), enabling holistic retrieval with seamless integration of new context.
Traceable & Explainable Retrieval
Explainable and traceable reasoning process, with each retrieved result containing exact page or section references.
No Time-to-First-Token Delay
Retrieval happens during generation time, allowing responses to stream immediately without waiting for a separate retrieval phase.
Lightweight Infra
Requires only a lightweight tree index (JSON) that integrates with mainstream databases. No extra infra needed.
Best for Domain-Specific Document Analysis
Financial reports and SEC filings
Regulatory and compliance documents
Healthcare and medical reports
Legal contracts and case law
Technical manuals and scientific documentation
Vector DB
Semantic Similarity
Low Retrieval Accuracy
Relies on semantic similarity, unreliable for domain-specific data where similarity does not imply relevance.
Context-Independent Retrieval
Embedding models lack the capacity to effectively incorporate chat context or specialized knowledge into retrieval, requiring fine-tuning to adapt to new context.
Black-box Retrieval without Traceability
Often lacks clear traceability to source documents, difficult to verify information or understand retrieval decisions.
Time-to-First-Token Delay
Retrieval is separate from generation, requiring users to wait for the entire retrieval phase to complete before the response begins streaming.
Extra Infra Overhead
Requires a separate embedding pipeline, vector database, and additional infra, with sync and maintenance overhead.
Best for Generic & Exploratory Applications
Vibe retrieval
Semantic recommendation systems
Creative writing and ideation tools
Short news/email retrieval
Generic knowledge question answering
Start building with PageIndex
01
Higher Accuracy
Relevance beyond similarity
02
Traceable & Explainable
Reasoning-driven retrieval with references
03
No Chunking
Preserves document structure
04
No Vector DB
No extra infra overhead
05
No Top-K
Retrieves all relevant passages
06
Context-aware Retrieval
Retrieval depends on full context
07
Human-like Retrieval
Mirrors how humans read
