The Process

Even the Experts Need a Flowchart

Hey, I'm Abhi. I frequently listen to pods like Hollinger & Duncan and The Lowe Post. One thing became ridiculously clear listening to the experts try to untangle the new CBA: its complicated. Like, "Wait, does the second apron prevent you from using the taxpayer mid-level exception only if you've also traded a future first-round pick seven years from now?" level of complicated.

It's a maze of cross-references, defined terms, and tables that can make anyone's head spin.

So, I got curious. Could I build an AI agent that doesn't just search the 696-page document, but actually understands it? An expert co-pilot that could answer a simple question like "How does the luxury tax work?" and not miss the five different sections that actually define it. With that foundation, I could build a tool that would help me answer questions about the CBA and help with more complex questions.

Teaching LLMs to Read a Legal Nightmare

Building this wasn't about just dumping a PDF into a chatbot. That fails instantly. The secret sauce is in how you prepare the data. I used a multi-level strategy designed specifically for dense, structured documents like the CBA.

1. The "Map & Zoom" Parsing Strategy

First, I didn't treat the whole document as one big blob of text. I broke it down into two types of "chunks":

Title Chunks (The Map): I took every high-level Article and Section from the Table of Contents and used a LLM to create a rich, keyword-heavy summary for each. This chunk knows, for example, that "Article VII, Section 2" is where you find rules about the "Salary Cap," "Luxury Tax," and the "Apron." It's the map of the entire document.

Content Chunks (The Details): I then took every single granular subsection—every (a), (i), and (1)—and turned it into its own tiny chunk. This is where the actual, specific rules live.

2. Smart Embeddings: Giving the Chunks Context

When I converted these chunks into vectors (the numbers that AI understands), I didn't just embed the raw text. I gave it context. A specific chunk about a tax rate isn't just a number; it's Article VII → Section 2 → Tax Level → Tax Bracket Table. By prepending this hierarchy to the text before embedding, the search becomes much smarter.

3. The Multi-Step RAG System: Mimicking a Real Expert

This is where it all comes together. When a user asks a question, our system works like a real expert would:

Step 1: Consult the Map (Broad Search): I do a vector search to find the most relevant high-level "Title" chunks. When you ask about "2nd apron restrictions," it immediately finds the enriched Article VII, Section 2 summary because it's packed with those keywords.

Step 2: Zoom In on the Details (Filtered Retrieval): The system sees that Article VII, Section 2 is the place to be. It then performs a second, targeted retrieval to pull all the detailed "Content" chunks from that specific section. This is how it finds the crucial "Transaction Restrictions Table" that a simple search would have missed.

Step 3: Synthesize the Answer: Finally, it hands all this rich, interconnected context—both the high-level summary and the specific rules—to the LLM (GPT-4o-mini) with a strict prompt: "You are an expert. Use only this context. Explain it simply and cite your sources."

This "map-then-zoom" approach ensures the AI has the complete picture, allowing it to connect a definition in one place to its consequences listed ten pages later, just like a lawyer would.