How It Works

The Big Picture

The Teton Scroll automatically collects thousands of government documents, extracts their content, and makes them searchable through AI. Here's how it works, from document discovery to answering your questions.

Step 1: Document Discovery

Our automated scrapers monitor government portals every day, looking for new documents:

CivicClerk — The county's meeting management system, where agendas and minutes are published
County websites — Board packets, resolutions, and other public records

We track documents from multiple boards:

Board of County Commissioners (BCC)
Planning & Zoning Commission (P&Z)
START Bus Transit
And more as we expand coverage

When a new document is discovered, it enters our processing pipeline.

Step 2: Processing Pipeline

Each document goes through several stages before becoming searchable:

Download & Storage — PDFs are fetched from government servers and stored securely. We keep track of when each document was published and which meeting it belongs to.

Text Extraction — PDFs are converted to searchable text. This preserves the content while making it accessible to our search system.

AI Analysis — This is where it gets interesting. Our AI reads each document and extracts:

People & Organizations — Who is mentioned? What roles do they play?
Events — Meeting dates, times, and locations
Votes & Decisions — Motions, how each commissioner voted, and outcomes
Glossary Terms — Acronyms and jargon with definitions (BCC, CUP, LDR, etc.)
Summaries — A concise overview of each document's content

Chunking — Documents are split into overlapping segments. This helps the search system find relevant passages without losing context at the edges.

Embedding — Each chunk is converted into a "vector embedding" — a numerical representation that captures its meaning. This enables semantic search: finding content by meaning, not just keywords.

Step 3: The Q&A System

When you ask a question, here's what happens behind the scenes:

Question Analysis — The system examines your question to understand what you're asking about:

Detects names of commissioners, staff, or organizations
Recognizes acronyms and government jargon
Identifies time periods ("recent," "last year," specific dates)
Classifies the topic (zoning, budget, housing, etc.)

Context Injection — Based on the analysis, relevant background information is provided to the AI:

Definitions for acronyms you mentioned
Profiles of people you asked about
Recent meeting summaries for time-based questions

Agentic Search — The AI (powered by Claude) doesn't just do a single search. It autonomously decides what to look for, refines its queries, and explores multiple angles:

Searches documents by content and keywords
Looks up entity profiles and voting records
Finds related meetings and decisions
Checks glossary definitions

Citation-Backed Answers — Every claim in the response is linked to a source document. You can click through to verify any fact yourself. If the AI can't find relevant information, it tells you rather than guessing.

Data Coverage

Speed — New documents typically appear within 24-48 hours of being published
Scope — Meeting minutes, agendas, board packets, resolutions, and public hearing records
History — We're continuously backfilling historical documents to expand coverage