🌲 Vectorless RAG for Apple SEC Filings

Tree-Based Reasoning Retrieval — No Vectors, No Embeddings, No Chunking

---

📋 Problem Statement

Traditional RAG (Retrieval-Augmented Generation) systems for financial documents face critical issues:

Problem	Impact
Chunking destroys context	A 100-page 10-K filing gets split into 500+ text chunks. Tables, cross-references, and section relationships are lost.
Embeddings miss nuance	Vector similarity search retrieves semantically similar text, not structurally relevant text. Asking "What is Apple's gross margin?" might retrieve a sentence mentioning margins rather than the actual financial table.
Top-K is noisy	Returning the top 5 most similar chunks often includes irrelevant content, forcing the LLM to sift through noise.
No transparency	Users can't see why a specific chunk was retrieved. The process is a black box.

The core question: Can we build a RAG system that retrieves information the way a human analyst would — by understanding the document's structure and reasoning about where the answer lives?

---

💡 Our Solution: Vectorless RAG with PageIndex

Instead of vectors + embeddings, we use tree-based reasoning retrieval:

Traditional RAG                          Vectorless RAG (This Project)
─────────────────                        ─────────────────────────────
Document → Chunk → Embed → Vector DB     Document → Parse → Tree Structure (JSON)
Query → Embed → Similarity Search         Query → LLM Reads Tree → Reasons About Sections
Top-K Chunks → LLM → Answer              Targeted Sections → LLM → Grounded Answer

How It Works

graph LR
    A["📄 SEC 10-K Filing<br/>(HTML)"] --> B["📝 Structured Markdown<br/>(Heading Hierarchy)"]
    B --> C["🌲 PageIndex Tree<br/>(JSON with Summaries)"]
    C --> D["🧠 Claude Reasons<br/>Over Tree Structure"]
    D --> E["📖 Fetch Targeted<br/>Sections Only"]
    E --> F["💡 Grounded Answer<br/>with Citations"]
    
    style A fill:#ef4444,color:#fff
    style C fill:#6366f1,color:#fff
    style D fill:#10b981,color:#fff
    style F fill:#22d3ee,color:#000

Loading

Step 1 — Tree Indexing: Each 10-K filing is converted into a hierarchical JSON tree (like an intelligent table of contents). Each node contains a title, summary, and full text content.

Step 2 — LLM Reasoning: When you ask a question, Claude reads the tree structure (titles + summaries only) and reasons about which sections are most likely to contain the answer — just like a human analyst scanning a table of contents.

Step 3 — Targeted Retrieval: Only the relevant sections are fetched (typically 1-3 out of 55 nodes), not a noisy top-K list.

Step 4 — Grounded Answer: The answer is generated solely from the retrieved sections, with precise citations to specific items and line numbers.

---

🎯 What We Built

Data Pipeline

• HTML → Markdown Converter (convert_html_to_md.py) — Transforms SEC EDGAR HTML filings into clean Markdown with strict heading hierarchy (# Title → ## PART → ### Item)
• Batch Tree Indexer (batch_index.py) — Generates PageIndex tree structures for all 6 filings (FY2020–FY2025)
• Query Pipeline (query_sec_filing.py) — 3-step reasoning → retrieval → synthesis pipeline

Web Interface

• Real-time Streaming UI — Watch the RAG pipeline work step-by-step via Server-Sent Events
• Interactive Tree Visualization — Explore the document hierarchy, see which nodes get selected
• Rate-Limited Demo — 5 free queries to try the system (protects API costs)

Live pipeline: Claude reasons over the tree, selects Node 0039 (MD&A) and Node 0041 (Financial Statements), then retrieves their full content

---

📊 Results

We tested the system across 6 Apple 10-K filings (FY2020–FY2025). Here are sample results:

Question	Filing	Answer	Sections Used	Time
What was Apple's total revenue?	FY2025	$416,161M (+6% YoY)	Item 7 — MD&A	11.8s
What is Apple's gross margin?	FY2025	$195,201M (46.9%) · Products: 36.8%, Services: 75.4%	Item 7 — MD&A	11.3s
What was R&D spending?	FY2025	$34.55B (+10% YoY from $31.37B)	Item 7 — MD&A	9.9s
How many employees?	FY2025	166,000 FTE	Item 1 — Business	9.9s
What are the main risk factors?	FY2020	COVID-19, supply chain, competition, regulation	Item 1A — Risk Factors	10.1s

Result: Precise gross margin breakdown with Products (36.8%) vs Services (75.4%) and YoY comparison — all cited from Item 7

Why Vectorless?

Metric	Traditional RAG	Vectorless RAG
Retrieval Precision	Retrieves 5-10 chunks, many irrelevant	Retrieves 1-3 exact sections
Context Preservation	Chunks lose table/section context	Full sections with all tables intact
Transparency	Black box — can't explain why chunks were retrieved	Full reasoning trace — see exactly why each section was selected
Setup Complexity	Vector DB + embeddings model + chunking strategy + reranker	Just a JSON tree + LLM
Cost	Embedding cost + vector DB hosting + LLM	LLM only

---

🚀 Quick Start

Prerequisites

• Python 3.9+
• Anthropic API Key

Installation

# Clone the repository
git clone https://github.com/nikhilreddy00/vectorless-rag.git
cd vectorless-rag

# Create virtual environment
python3 -m venv venv
source venv/bin/activate

# Install dependencies
pip install flask litellm anthropic python-dotenv beautifulsoup4 markdownify

# Set your API key
echo 'ANTHROPIC_API_KEY=your-key-here' > .env

Run the Web Interface

python3 app.py
# Open http://localhost:5001

Run from Command Line

python3 query_sec_filing.py
# Select a filing → Ask any question

Re-index Filings (Optional)

# Convert HTML to Markdown
python3 convert_html_to_md.py

# Generate tree structures
python3 batch_index.py

---

📁 Project Structure

vectorless-rag/
│
├── app.py                      ← 🌐 Flask web server with SSE streaming
├── query_sec_filing.py         ← 📟 CLI query pipeline
├── convert_html_to_md.py       ← 🔄 HTML → structured Markdown converter
├── batch_index.py              ← 🌲 Batch tree generation for all filings
│
├── static/
│   ├── index.html              ← Frontend HTML
│   ├── style.css               ← Dark theme design system
│   └── app.js                  ← Frontend logic (tree viz, SSE, pipeline)
│
├── documents/
│   ├── aapl-20200926.html      ← Original SEC filings (HTML from EDGAR)
│   ├── ... (6 files)
│   └── markdown/               ← Converted Markdown files
│       ├── aapl-20200926.md
│       └── ... (6 files)
│
├── results/                    ← 🌲 Generated tree structures (JSON)
│   ├── aapl-20200926_structure.json  (51 nodes)
│   ├── aapl-20210925_structure.json  (53 nodes)
│   ├── aapl-20220924_structure.json  (53 nodes)
│   ├── aapl-20230930_structure.json  (55 nodes)
│   ├── aapl-20240928_structure.json  (55 nodes)
│   └── aapl-20250927_structure.json  (55 nodes)
│
├── assets/                     ← Screenshots for README
├── PageIndex/                  ← PageIndex framework (submodule)
├── .env                        ← API key (not committed)
└── usage.json                  ← Rate limit counter

---

🔧 Tech Stack

Component	Technology
Tree Generation	PageIndex
LLM	Anthropic Claude Sonnet (via LiteLLM)
Backend	Flask + Server-Sent Events
Frontend	Vanilla HTML/CSS/JS (dark theme, no frameworks)
Data Source	SEC EDGAR (Apple 10-K filings, FY2020–FY2025)

---

⚠️ Demo Rate Limit

The live demo is limited to 5 queries to protect API costs. After 5 queries, the interface will show a friendly limit message. To run unlimited queries, clone the repo and use your own Anthropic API key.

To reset the counter:

rm usage.json

---

📜 License

This project is for educational and research purposes. SEC filing data is publicly available from SEC EDGAR.

---

Built with 🌲 PageIndex + 🤖 Claude · No vectors were harmed in the making of this project