Vectorless RAG — Web Interface

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

Vectorless RAG Claude Sonnet SEC 10-K PageIndex Python 3.9+

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

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?

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

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

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.

• 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

• 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 RAG Pipeline — Reasoning, Node Selection, and Retrieval
Live pipeline: Claude reasons over the tree, selects Node 0039 (MD&A) and Node 0041 (Financial Statements), then retrieves their full content

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

Query Result — Gross Margin with precise financial data and citations
Result: Precise gross margin breakdown with Products (36.8%) vs Services (75.4%) and YoY comparison — all cited from Item 7

• Python 3.9+
• Anthropic API Key

# 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

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

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

# Convert HTML to Markdown
python3 convert_html_to_md.py
# Generate tree structures
python3 batch_index.py

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

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

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