Help Center

GraphRAG replaces traditional vector-only RAG with a 5-layer hybrid retrieval network: Layer 1 (Graph Traversals across entity nodes like People, Companies, Technologies, Documents), Layer 2 (Pinecone 1536-dim Vector Search), Layer 3 (Firestore Metadata Search), Layer 4 (Long-Term Memory), and Layer 5 (Reciprocal Rank Fusion context compression).

Knowledge Studio allows universal source ingestion (PDF, Word, Excel, PPTX, CSV, JSON, Markdown, YouTube transcripts, web articles, Whisper audio). It automatically extracts entity-relationship subgraphs and supports two-host AI Audio Overviews (podcasts), sandboxed Python/SQL Code Notebooks, mind maps, presentation decks, flashcards, and quizzes.

Every incoming prompt is evaluated by our Supervisor Router Agent built on LangGraph. The Supervisor analyzes intent, complexity, and domain context, routing execution across 15 specialized domain agent nodes (Coding, Science, Writing, Finance, Legal, Medical, Math, Business, Data Science, Debate, etc.).

iHailey maintains persistent context across 4 distinct memory layers: L1 Working Memory (active LangGraph execution state), L2 Redis Cache (microsecond retrieval), L3 Firestore Memory Store (persistent structured facts & entity graphs), and L4 Pinecone Vector Memory (1536-dimensional semantic embeddings).

Audio Overviews generate engaging podcast discussions between two AI hosts analyzing your uploaded Knowledge Studio documents. Powered by OpenAI Text-to-Speech (tts-1) with beginner, academic, and executive briefing modes.

Code Notebooks execute Python and SQL code in a sandboxed analytical environment, producing interactive charts, tables, and statistics directly from document data.