cognis

LLM providers, chains, agents, memory, and tools for Rust.

Workspace | API Docs

cognis is the implementation layer of the Cognis framework. It provides concrete LLM provider integrations, agent execution, composable chains, conversation memory, document processing, and built-in tools — all behind feature flags so you only compile what you use.

Quick Start

[dependencies]
cognis = { version = "0.1", features = ["anthropic"] }
cognis-core = "0.1"
tokio = { version = "1", features = ["full"] }

use cognis::chat_models::anthropic::ChatAnthropic;
use cognis_core::runnables::Runnable;
use serde_json::json;

let model = ChatAnthropic::new("claude-sonnet-4-20250514");
let result = model.invoke(json!({"messages": []}), None).await?;

LLM Providers

Enable only what you need:

cognis = { version = "0.1", features = ["openai", "anthropic"] }  # pick providers
cognis = { version = "0.1", features = ["all-providers"] }         # or grab everything

What's Inside

Agents — Executor with a pluggable middleware pipeline: retry, PII redaction, tool selection, human-in-the-loop, summarization, and more.

Chains — LLM chain, conversation chain, sequential chain, extraction chain, structured output chain, router chain, retrieval QA chain.

Memory — Buffer, window, summary, and entity memory strategies for multi-turn conversations.

Document Loaders — Text, CSV, JSON, and directory loaders for ingesting data.

Text Splitters — Character, recursive, markdown, HTML, JSON, code, and token-aware splitters for chunking documents.

Embeddings — OpenAI and Ollama embedding providers with batch support.

Retrievers — Caching, reranking, query translation, and multi-query retrievers.

Tools — Calculator, shell command, and JSON query tools out of the box.

Additional Feature Flags

Part of the Cognis Workspace