LoreGraph

Evidence-backed repository intelligence for VS Code.

LoreGraph extracts the business concepts, configuration dependencies, and code relationships hidden inside complex enterprise codebases, builds them into a knowledge graph, and turns that graph into an interactive visual map and Docusaurus documentation — without ever inventing a fact.

Core thesis: No evidence, no documentation claim. LoreGraph is not a generic AI doc generator. Every entity, relationship, and sentence it produces links back to a specific file and line in your repository.

The problem: tribal knowledge

Large enterprise systems bury years of undocumented knowledge in config files, environment variables, feature flags, Spring classes, Kafka topics, database schemas, comments, acronyms, and half-stale docs. New engineers can't tell what a business term means, where it's implemented, which configs control it, or which docs to trust.

LoreGraph makes that knowledge explicit and auditable.

How it works

Repository Scanner → Config Parser → Code Usage Resolver → Domain Term Extractor
→ Annotation Extractor → Relationship Builder → Knowledge Graph → Confidence Scorer
→ Visual Graph Builder → Documentation Generator → Drift Detector

1. Scan the workspace (skipping node_modules, target, .env, secrets, etc.).
2. Parse code with a tree-sitter AST — Java & TypeScript/JS/JSX get framework-aware extraction (Spring, Nest, Kafka, REST), while Python, Go, C and C++ get structural extraction (classes/structs/interfaces, functions, inheritance, references); config/SQL/markdown use lightweight parsers. All scope-aware and import-resolved, never from comments or strings.
3. Build a knowledge graph where code symbols are identified by file + fully-qualified name (so same-named classes don't collapse) and every node and edge carries SourceRef evidence.
4. Score confidence: human annotations = high, multi-source = boosted, naming-only = low.
5. Trace multi-hop flows by traversing the graph (event: producer → topic → consumer; request: endpoint → controller → service → table).
6. Visualize the graph inside VS Code (React + React Flow) with an edge-aware layered layout (dagre).
7. Generate Docusaurus docs from graph-backed claims only, with C4-style Mermaid + sequence diagrams.
8. Flag missing/uncertain knowledge as explicit open questions.

Anti-hallucination design

• Documentation claims are generated only from extracted graph evidence.
• Inferred meanings are hedged ("appears to", "is likely related to").
• Insufficient evidence produces an open question, never an invented answer.
• Every generated page has an Evidence section and a confidence badge.
• Existing docs/comments and multiple independent sources raise confidence; naming alone stays low.
• Human @lore annotations produce high-confidence, verified claims.

Human-in-the-loop annotations

Drop @lore comments anywhere (//, #, *, --):

// @lore.term Reconciliation
// @lore.purpose Compares trade records across systems to identify breaks.
// @lore.owner Trade Operations Platform
// @lore.risk Changing retry settings may delay end-of-day break detection.

These become verified, high-confidence graph claims.

Visual Graph Explorer

Four modes, one interactive canvas:

• Concept Graph — Obsidian-style map of business terms ↔ configs, services, topics, tables, docs.
• Config Dependency Map — where each config key is defined, read, and documented.
• Architecture Map — simplified C4 view: API → services → messaging → data & config.
• Flow Map — directional event flows from Kafka producers to consumers.

Search, filter by entity type (click the legend) and confidence, focus a node, click evidence to jump to source, and inspect any node's claims, related entities, and open questions.

Generated documentation

LoreGraph: Generate Documentation Site writes a ready-to-build Docusaurus project:

loregraph-docs/
  docs/
    intro.md
    onboarding/{start-here,knowledge-map}.md
    business-terms/{glossary, <term>.md ...}
    configuration/{overview,config-keys,feature-flags,environment-variables}.md
    architecture/{overview,service-map,concept-map}.md
    flows/{inferred-flows,kafka-flows}.md
    operations/{open-questions,docs-drift-report}.md
  docusaurus.config.ts
  sidebars.ts
  package.json

Every page includes title, status, confidence, summary, evidence-backed claims, related concepts/configs/code, open questions, and a "Needs Human Review" section where applicable.

Commands (Command Palette → "LoreGraph:")

Configuration

Local-first, no external API

LoreGraph runs entirely offline. The LLMProvider interface exists for future AI providers, but the default LocalTemplateProvider generates all documentation deterministically from the grounded EvidenceBackedPrompt — which, by contract, only ever exposes graph evidence, never raw source.

Roadmap

• Real LLM provider integration (grounded by the same evidence contract)
• GitHub PR documentation drift checks
• Jira/Confluence import
• CODEOWNERS-based ownership mapping (partial today)
• Cross-file TypeScript reference resolution via module-path imports (AST today resolves by import/short name)
• Docusaurus deployment workflow
• Team review workflow for verifying generated claims
• Semantic embeddings for better term clustering
• Call-graph-precise request flows (method-level call edges, building on the current reference graph)

License

MIT