Zerda documentation

Retrieval-augmented generation engine plus a library of RAG strategies (agentic, graph, multi-hop, streaming and more).

Multi-tier memory architecture for conversational agents: buffers, summaries, entities, semantic recall and durable storage.

A hierarchical, hybrid-scoring query router for LLM applications.

An intelligent, multi-tenant, cost-aware semantic caching module for LLM applications.

Production-grade RAG chunking for Arabic, multilingual and English text.

A dependency-light unified interface for calling LLM providers, with tool calling across all seven backends.

Unified text-embedding interface — with first-class attention to Arabic — across local and hosted backends.

A unified interface over three different vector-search backends.

Retrieval-orchestration and context-engineering for RAG, with support for mixed Arabic/English content.

Converts files, web pages, directories, knowledge graphs and tables into a normalized LoadedDocument.

Builds AI-ready, context-aware prompts for RAG — guardrailed, token-budgeted and provider-ready.

Output parsing, validation, fault-tolerance and retry engine for LLM/RAG/agent pipelines.

A generic, backend-agnostic tracing system for Zerda.

Build, discover, secure and orchestrate plugins that an LLM can call.