Factguard News Fact Checking

I built FactGuard, a Python system for real-time news analysis and fact-checking. It combines news scraping, claim extraction, evidence-backed verdicts, agent orchestration, graph memory, vector retrieval, credibility tracking, prediction, benchmark evaluation, and a Streamlit dashboard. The repository is on my GitHub as @abdulwasaeee; it started and was last updated in 2026-04.

FactGuard uses a multi-agent LangGraph pipeline, Neo4j graph memory, and ChromaDB vector retrieval to produce evidence-backed verdicts.

The system scrapes news from RSS feeds, NewsAPI/Tavily, and Telegram, then extracts claims, named entities, and image captions from articles.

For each claim, the fact-checking flow runs cache check, freshness gate, live web search, RAG retrieval, and verdict synthesis; multi-agent debate and cross-modal image-consistency checks are optional.

FactGuard tracks entity credibility over time using Neo4j snapshots and predicts future credibility changes using linear-regression trend detection.

The orchestration layer contains 13 graph node functions.

The ChromaDB VectorStore has 5 collections.

The code is organized around agents for scraping, preprocessing, fact-checking, memory, source reflection, entity tracking, and prediction. Memory provides a unified Neo4j and ChromaDB facade; the orchestration package builds the LangGraph StateGraph, defines its nodes and conditional routing, and maintains FactCheckState. Tools cover ChromaDB retrieval with RRF fusion, Tavily search, freshness checks, image–claim comparison with SigLIP/CLIP, and reranking.

The data flow sends scraped RawArticle items through preprocessing, which yields a Source, Article, claims, and image caption for storage in Neo4j and ChromaDB. Fact checking retrieves similar claims, reads and writes verdicts, consults source credibility, and can search the live web; its output contains a verdict, confidence, evidence links, and bias. Entity tracking stores credibility snapshots, prediction resolves and records predictions, and the Streamlit frontend exposes live verdicts, credibility charts, and prediction tables.

The debate confidence threshold defaults to 70 on a 0–100 scale; debate is configured to run below that threshold.

• The default feature flags are USE_GRAPH_RAG=false, USE_DEBATE=false, USE_CROSS_MODAL=false, USE_RERANKER=false, OFFLINE_MODE=false, and DRY_RUN=false. Graph RAG adds Neo4j entity-graph context; offline mode skips Tavily live search, and dry-run mode skips database writes.

FactGuard evaluates verdict classification on LIAR and Wikipedia fact-checking on FEVER using Macro-F1, Entity-F1, and Precision@5; Factify evaluates multimodal claims and images using Macro-F1 and Image-Mismatch F1.

The evaluation script loads real datasets through HuggingFace datasets and falls back to synthetic mock data when that package or the real datasets are unavailable.

My team divided the work across three tasks: data and memory, core agentic AI, and full-stack development and evaluation. The captured README does not report measured benchmark scores, latency, cost, or production reliability.

I need Python 3.11 or later and Docker for Neo4j and ChromaDB.

docker-compose up -d
pip install -r requirements.txt
python -m spacy download en_core_web_sm
cp .env.example .env
# Edit .env — at minimum set OPENAI_API_KEY and TAVILY_API_KEY
streamlit run frontend/app.py

The local services use Neo4j at bolt://localhost:7687 (browser at http://localhost:7474) and ChromaDB at http://localhost:8000; the Streamlit dashboard is at http://localhost:8501.

python -c "from agents.fact_check_agent import fact_check_claim; print(fact_check_claim('Tesla recalled 500k cars'))"
python -m agents.entity_tracker    # defaults to "Tesla", last 24h
python -m agents.prediction_agent  # defaults to "Tesla"
python -m evaluation.evaluation --dataset liar --samples 50
python -m evaluation.evaluation --dataset all  --samples 100

• The README assigns Task 1, Data & Memory Engineer, to branch data_memory_dev; Task 2, Core Agentic AI Engineer, to branch main (FakeNewsAgent); and Task 3, Full-Stack & Evaluation Engineer, to branch task3.

The captured project description leaves implementation details open, including verdict labels, confidence calculation, cache policy, freshness criteria, retrieval settings, database schema and collection names, prediction rules, and how debate outputs are combined. It also does not identify the default news feeds or explain the newsapi.py fetcher’s Tavily naming.