back

Time: 7 minute read

Created: August 5, 2024

Author: Justin Torre

The Emerging LLM Stack: A New Paradigm in Tech Architecture

See companion blog here before reading this.

LLM stack guide

As seasoned developers, we've witnessed the ebb and flow of numerous tech stacks. From the meteoric rise of MERN to the gradual decline of AngularJS, and the innovative emergence of Jamstack, tech stacks have been the backbone of web development evolution.

Enter the LLM Stack - a cutting-edge tech stack designed to revolutionize how developers build and scale LLM (Large Language Model) applications.

Why a New Stack for LLM Applications?

LLM applications are deceptively simple to kickstart, but scaling them unveils a Pandora's box of challenges:

  1. Platform Limitations: Traditional stacks struggle with the unique demands of LLM apps.
  2. Tooling Gaps: Existing tools often fall short in managing LLM-specific workflows.
  3. Observability Hurdles: Monitoring LLM performance requires specialized solutions.
  4. Security Concerns: LLMs introduce new vectors for data breaches and prompt injections.

To illustrate this evolution, we've created a companion blog that walks through a real-world LLM application's journey from MVP to scalable product.

You might be interested in:

Anatomy of the LLM Stack

Let's dissect the architecture of a robust LLM stack, based on our detailed example:

LLM Stack Architecture

Note: that this article is not going to tell you what stack to choose, but rather outlines the components for building an LLM stack.

Key Components

1. Observability Layer

This layer focuses on monitoring, tracking, and analyzing the performance and behavior of LLM applications. It provides insights into how the models are being used, their performance metrics, and helps in debugging and optimizing the system.

Observability Layer

ServiceCutting-Edge Solutions
ObservabilityHelicone, LangSmith, LangFuse, Traceloop, PromptLayer, HoneyHive, AgentOps
ClientsLangChain, LiteLLM, LlamaIndex
AgentsCrewAI, AutoGPT, DifyAI, Autogen
Prompting LayerHelicone Prompting, PromptLayer, PromptFoo
IntegrationsZapier, LlamaIndex

You might be interested:

2. Inference Layer

The Inference Layer is responsible for processing requests to the LLM, managing model deployments, and load balancing. It handles the actual execution of the language models and ensures efficient and scalable access to these models.

Inference Layer

ServiceInnovative Products
GatewayHelicone Gateway, Cloudflare AI, PortKey, KeywordsAI
LLM Load BalancerMartian, LiteLLM
Model ProvidersOpenAI (GPT), Meta (Llama), Google (Gemini), Mistral, Anthropic (Claude), TogetherAI, Anyscale, Groq, Amazon (Bedrock)

3. Testing & Experimentation Layer

This layer enables developers to test, evaluate, and experiment with different prompts, models, and configurations. It provides tools for fine-tuning models, running experiments, and assessing the quality and performance of LLM outputs.

Testing & Experimentation Layer

ServiceNext-Gen Tools
DatastoresHelicone API, LangFuse, LangSmith
ExperimentationHelicone Experiments, Promptfoo, Braintrust
EvaluatorsPromptFoo, Lastmile, BaseRun
Fine TuningOpenPipe, Autonomi

Helicone: Your Gateway to LLM Excellence

Helicone isn't just another tool in the LLM stack - it's a game-changer. Our primary focus areas, Gateway and Observability, are crucial for building robust, scalable LLM applications.

Helicone in the LLM Stack

By integrating Helicone into your LLM workflow, you're not just adopting a tool; you're embracing a philosophy of efficiency, scalability, and deep insights into your LLM applications.

Ready to supercharge your LLM development? Dive deeper into Helicone's capabilities and see how it can transform your LLM stack today!