Features¶

WikiMind transforms raw information into a structured, interconnected knowledge base. Here is what it can do today.

Source Ingestion¶

Feed WikiMind from multiple source types:

• Web URLs -- Articles, blog posts, documentation pages. Extracted via trafilatura with full HTML cleanup.
• PDF documents -- Research papers, reports, slide decks. Processed by docling-serve for structured extraction (heading hierarchy, tables, OCR). Falls back to pymupdf when docling-serve is unavailable.
• Raw text -- Paste notes, meeting transcripts, or any plain text directly.
• YouTube videos -- Extracts the transcript automatically via youtube-transcript-api.

Sources are deduplicated by content hash, so re-ingesting the same document is a no-op.

LLM Compilation¶

Every ingested source is compiled into a structured wiki article by an LLM. The compilation produces:

• Title and summary -- Concise, specific article title and a two-sentence summary
• Key claims -- Specific, falsifiable claims extracted from the source, each tagged with a confidence level (sourced, inferred, or opinion) and optional direct quote
• Concepts -- Topic tags that connect the article to the broader knowledge graph
• Backlink suggestions -- Related articles with typed relationships (references, extends, supersedes)
• Open questions -- Gaps the source raises but does not answer
• Article body -- Full markdown article (300+ words) with analysis

Large documents (over 80K tokens) are automatically chunked and compiled in parts, then merged into a single article.

Multi-Provider LLM Router¶

WikiMind supports multiple LLM providers with automatic fallback:

Providers are auto-enabled when their API key is detected -- no manual configuration flags needed. The router falls back to the next available provider when one fails.

Monthly budget tracking with WebSocket alerts prevents surprise bills.

Conversational Q&A¶

Ask questions against your compiled wiki:

• Multi-turn conversations -- Follow-up questions carry context from prior turns
• Source citations -- Every answer cites the wiki articles it drew from
• Confidence scoring -- Answers are rated high/medium/low confidence
• File-back -- High-confidence answers are filed back into the wiki as new articles, making the wiki smarter over time
• Conversation forking -- Branch a conversation at any turn to explore a different line of reasoning
• Streaming -- Token-by-token streaming via Server-Sent Events for responsive UI

Knowledge Graph¶

Articles are interconnected through:

• Concepts -- Hierarchical topic taxonomy, auto-generated and LLM-rebuilt
• Backlinks -- Typed relationships between articles (references, extends, supersedes, contradicts)
• Concept pages -- Auto-generated articles that synthesize all sources tagged with a concept

Wiki Health & Linting¶

A built-in linter audits the wiki for quality issues:

• Contradiction detection -- Finds conflicting claims across articles using batched LLM analysis
• Orphan detection -- Identifies articles with no backlinks
• Health reports -- Scored reports with actionable findings
• Contradiction resolution -- Mark contradictions as resolved with a note

Authentication¶

Optional multi-user mode with OAuth2:

• Google and GitHub OAuth2 sign-in
• JWT sessions via HttpOnly BFF cookies
• Per-user data isolation (each user sees only their own sources, articles, and conversations)
• When disabled (default), WikiMind runs in single-user mode with no login required

Real-Time Updates¶

WebSocket events push live updates to the UI:

• Compilation progress and completion
• Budget warnings and alerts
• Linter findings
• Job status changes