PageIndex
Reasoning-based RAG for complex, long-form documents. Unlike vector search, PageIndex builds a hierarchical tree index and uses LLM reasoning to navigate it — delivering significantly better accuracy on financial reports, legal filings, technical manuals, and research papers.
Uses the PageIndex cloud API — no vector database or embedding pipeline needed.
Quick Start
import { Agent, openai, PageIndexToolkit } from "@radaros/core";
const pageindex = new PageIndexToolkit({
apiKey: process.env.PAGEINDEX_API_KEY,
});
const agent = new Agent({
name: "document-analyst",
model: openai("gpt-4o"),
instructions: "Analyze uploaded documents. Answer questions accurately with citations.",
tools: [...pageindex.getTools()],
});
const result = await agent.run(
"Submit https://example.com/annual-report.pdf and then summarize the revenue breakdown by segment."
);
Config
PageIndex API key. Falls back to PAGEINDEX_API_KEY env var. Get yours at dash.pageindex.ai. apiBase
string
default:"https://api.pageindex.ai"
API base URL. Override for self-hosted PageIndex deployments.
Request timeout in milliseconds. PDF processing can take time — the default is 2 minutes.
Max response characters returned per tool call.
| Tool | Description |
|---|
pageindex_submit | Submit a PDF document for tree indexing. Returns a doc_id for subsequent operations. |
pageindex_status | Check document processing status — returns tree structure when complete. |
pageindex_tree | Get the hierarchical tree structure of a processed document (semantic table of contents). |
pageindex_list | List all documents with IDs, names, statuses, and page counts. |
pageindex_chat | Ask questions about documents using reasoning-based RAG with optional citations. |
pageindex_retrieve | Retrieve specific sections from a document using tree-based search. |
pageindex_delete | Delete a document and all associated data. |
How It Works
PageIndex takes a fundamentally different approach from traditional vector RAG:
- Tree Indexing — Documents are parsed into a hierarchical tree of sections, subsections, and paragraphs with summaries at each level
- LLM Tree Search — At query time, an LLM navigates the tree from root to relevant leaves, using reasoning instead of embedding similarity
- No Vectors Needed — No embedding model, no vector database, no chunking strategy to tune
This approach excels on documents where structure matters: financial reports with complex tables, legal contracts with nested clauses, and technical specs with cross-references.
Use Cases
Document Q&A
// After submitting a document
await agent.run("What were the total operating expenses in Q3 2024?");
Multi-Document Analysis
await agent.run(
"Compare the risk factors section between the 2023 and 2024 annual reports."
);
const agent = new Agent({
name: "extractor",
model: openai("gpt-4o"),
instructions: "Extract structured data from documents with page citations.",
tools: [...pageindex.getTools()],
outputType: z.object({
items: z.array(z.object({
field: z.string(),
value: z.string(),
page: z.number(),
})),
}),
});
Environment Variables
| Variable | Description |
|---|
PAGEINDEX_API_KEY | PageIndex API key from dash.pageindex.ai |
Combining with RadarOS Knowledge
PageIndex works best for complex professional documents. For simpler content or when you need a fully local pipeline, combine it with RadarOS’s built-in vector knowledge base:
import { Agent, openai, PageIndexToolkit, InMemoryKnowledge } from "@radaros/core";
const agent = new Agent({
name: "hybrid-knowledge",
model: openai("gpt-4o"),
tools: [...new PageIndexToolkit({ apiKey: "..." }).getTools()],
knowledge: new InMemoryKnowledge({ /* local vector search for quick lookups */ }),
instructions: "Use PageIndex for complex document analysis. Use knowledge search for quick factual lookups.",
});
PageIndex is ideal for complex, structured documents (100+ pages). For short text snippets and FAQ-style retrieval, the built-in vector knowledge base is faster and cheaper.
