An LLM Wiki Engine That Fact-Checks Itself

What it does

Synthadoc ingests raw files—PDFs, spreadsheets, slides, web pages, even videos—and uses an LLM to synthesize them into a persistent wiki written as local Markdown. Instead of retrieving chunks at query time, it compiles cross-linked pages during ingestion, detects contradictions between sources, and annotates every substantive claim with a citation pointing to the exact line range in the original file.

The interesting bit

The project treats the wiki as the primary artifact, not a secondary view over a vector database. A built-in adversarial lint pass dispatches a second LLM to play devil’s advocate against every page, flagging overconfident claims and unsupported superlatives before a human ever reads them.

Key highlights

• Ingest-time compilation produces [[wikilinks]] and surfaces contradictions automatically, preserving both conflicting claims with citations.
• Claim-level provenance uses ^[filename:L-L] chips that jump to the exact source lines; broken citations are caught by the lint system.
• Every page moves through a five-state lifecycle—draft, active, contradicted, stale, archived—with a full audit trail of who changed what and why.
• Output is plain Markdown, so the wiki is browsable offline in Obsidian or any wiki-compliant editor without the engine running.
• Supports multiple LLM backends, including local Ollama models and free-tier APIs, keeping the stack local-first and subscription-agnostic.

Verdict

Worth a look if you manage a growing corpus that needs structure, cross-references, and source accountability. Skip it if you only need occasional ad-hoc answers over a handful of files and don’t want a maintained wiki artifact.