I started the Fact‑Lens repository in April 2026 and last updated it in April 2026.
The codebase is written in Python.
Two main pipeline entry points are pipeline.claim_pipeline.ClaimPipeline and pipeline.document_pipeline.DocumentPipeline.
The locked default runtime has the LLM verifier turned ON (ENABLE_LLM_VERIFIER=1).
Search provider configuration enables DuckDuckGo (DDG) and disables Tavily and SerpAPI.
I built FactLens, a Python repository on my GitHub, to tackle citation hallucination in single‑prompt LLM generation and naïve RAG by using an adversarial multi‑agent system on LangGraph.
FactLens orchestrates 12 specialized agents organized across distinct evaluation phases.
Three Parallel Retrieval Channels (RunnableParallel) are used to run the Web Scout, Archive Hunter, and Data Extractor agents simultaneously.
Retrieval priority waterfall comprises seven ordered sources.
The system follows five phases—Ingestion & Decomposition, Parallel Multi‑Channel Retrieval, Cross‑Channel Synthesis & Quality Audit, Adversarial Audit & Rebuttal, and Consensus Arbitration & Provenance—storing full provenance in SQLite.