Token usage statistics from LLM response.

A text passage for HippoRAG 2 indexing. Follows HippoRAG 2's corpus format: {"title": "...", "text": "...", "idx": 0} We extend

Retriever wrapper for sentence-transformers models (e.g. all-MiniLM-L6-v2).

Builds HippoRAG 2 corpus from ATM-Bench data sources.

Main class for processing videos and extracting comprehensive metadata.

Batch processor for large image collections with resume capability and concurrent processing.

Batch processor for large video collections with resume capability and concurrent processing.

Batch processor for large email collections with resume capability and concurrent processing.

Main class for processing emails and extracting comprehensive metadata.

A corpus of text passages for HippoRAG 2.

Wrapper that adapts LocalHuggingFaceEmbedder to HippoRAG's embedding interface. HippoRAG expects embedding models with specific method signa

Hybrid BM25 + Embedding retriever matching original A-Mem paper. Combines lexical (BM25) and semantic (embedding) scores with configurable weight

Main class for processing images and extracting comprehensive metadata.

Local HuggingFace embedding model for HippoRAG 2. This class provides the same interface as HippoRAG's embedding models but runs locally

Track processing progress and estimate completion time.

Track processing progress with timing information.

Track processing progress and estimate completion time.

Encode image and text metadata as *separate* embeddings, then concatenate. This addresses the concern that image tokens may dominate the embeddin