Training Sources

ConCRG learns your application through four complementary sources. They run in parallel during training and each contributes a different angle of knowledge.

---

Overview

Source	Input	What It Learns	LLM Used?
Probe	Live running app	Pages, navigation, UI elements, workflows	Yes — screen analysis
Code	TypeScript source	Routes, data types, permissions, API contracts	Hybrid — AST + LLM
Docs	Documentation URL	Concepts, feature descriptions, workflows	Yes — per-page extraction
Chat	Conversation with PM/developer	Domain rules, business logic, edge cases	Yes — conversational

You can run all four together or pick a subset. Each source produces RDF triples that are merged into the shared knowledge store.

---

Probe

The Probe is ConCRG's autonomous DOM explorer. It runs inside the host app's browser context — not in a separate Playwright session — so it sees the actual rendered UI with real authentication and data.

How it works:

1. Starts at the app's root route
2. Serializes the DOM tree (labels, roles, nav links, interactive elements)
3. Captures a screenshot
4. Sends the DOM snapshot + screenshot to Claude for analysis
5. Extracts page understanding, elements, workflows, and triples
6. Follows navigation links and repeats for each discovered route

Crawler modes:

• legacy — Breadth-first queue walker
• stategraph — Models the app as a state machine; discovers transitions, modal states, and conditional UI

Configuration:

crawlerMode: 'stategraph',
crawler: {
  maxNodes: 50,
  maxActionsPerNode: 10,
  maxInstancesPerRoutePattern: 3,
}

Deep dive: Probe training →

---

Code

The Code source analyzes your TypeScript source files to extract structural knowledge that's hard to observe from the DOM alone — route definitions, interface shapes, permission guards, and API contracts.

What it extracts:

• Route declarations and their corresponding components
• TypeScript interfaces (data models)
• JSX structure and prop relationships
• API endpoint handlers and their contracts
• Permission/role guards on routes and components

Two modes:

• Deterministic — TypeScript AST parsing (fast, precise)
• LLM-enhanced — Claude interprets complex handlers and business logic (useAgents: true)

Setup:

const config = {
  sourceCodePath: '/absolute/path/to/your/src',
  useAgentsV2: true,   // recommended for complex codebases
};

Or upload a zip file via the training panel for remote analysis.

Deep dive: Code analysis →

---

Docs

The Docs source crawls your existing documentation website and extracts product knowledge from it. Useful for bridging the gap between what your app can do (Probe + Code) and the business-level concepts behind it (Docs).

How it works:

1. Starts at the docsUrl you configure
2. Discovers all linked pages within the same domain
3. Extracts article content using @mozilla/readability
4. Converts HTML to Markdown
5. Sends each page to Claude for knowledge extraction (batched)
6. Streams progress in real time as pages are processed

Setup:

const config = {
  docsUrl: 'https://docs.yourproduct.com',
};

Streaming response: The docs crawl streams NDJSON progress updates so you can see pages being processed in real time.

Deep dive: Docs crawling →

---

Chat

The Chat source lets you teach ConCRG directly through conversation. This is ideal for domain knowledge that isn't visible in the UI or documented anywhere — business rules, edge cases, terminology, "why does this work this way?" context.

Example conversation:

You: In our system, a "settlement" only applies to bond trades,
     not equity trades. The UI looks the same but the validation
     rules are completely different.

ConCRG: Got it. I'll record that settlement workflows for bonds
         have different validation from equity. Can you tell me
         what the key differences are?

Each exchange is converted into knowledge triples and added to the store.

Deep dive: Chat training →

---

Combining Sources

The four sources are complementary, not redundant:

• Probe discovers what exists and how to navigate it
• Code reveals the underlying data model and permissions
• Docs provides the business-level concepts and terminology
• Chat adds tacit knowledge that no automated source can capture

The more sources you use, the richer and more accurate the knowledge graph becomes.