What it does AnyQ is a configurable, plugin-based FAQ question-answering system. You feed it a set of question-answer pairs, and it handles the rest: analyzing the incoming question, retrieving candidates via inverted index or semantic search, matching with both lexical and neural features, and re-ranking. It ships with 20+ plugins covering Chinese word segmentation, Solr-based retrieval, BM25/Jaccard/Cosine matching, and SimNet neural semantic matching.

The interesting bit The real payload is SimNet, Baidu’s in-house semantic matching framework (circa 2013) that runs on PaddleFluid and TensorFlow. It bundles both representation-based models (BOW, CNN, RNN) and interaction-based models (MatchPyramid, MV-LSTM, K-NRM). Baidu claims their pre-trained SimNet-BOW model beats literal similarity methods by 5%+ AUC on real FAQ data — though they don’t specify which datasets.

Key highlights

• Modular pipeline: Question Analysis → Retrieval → Matching → Re-Rank, all swappable via config
• Dual retrieval: classical Solr inverted index + ANNOY-based approximate nearest neighbor semantic search
• SimNet supports both PaddleFluid and TensorFlow backends with extensible model definitions
• HTTP server and C++ library interfaces both included
• Docker images provided for macOS/Windows; Linux builds need cmake 3.0+, g++ ≥4.8.2, bison ≥3.0

Caveats

• Documentation is primarily in Chinese; English README exists but may lag
• Pre-trained model claims lack reproducible benchmarks or dataset names
• Build dependencies are dated (cmake 3.2.2 “recommended”, g++ 4.8.2 minimum)

Verdict Worth a look if you’re building Chinese-language FAQ bots and want a battle-tested retrieval-plus-neural pipeline without writing glue code. Skip if you need modern transformer-based models or extensive English-language support — the ecosystem here is firmly 2013-2018 era deep learning.