Use cases

From real-time chatbots to batch pipelines processing millions of tokens — Meridian is built for every workload. Each pattern ships with a model recommendation tuned for cost, latency, and accuracy.

Chatbot

Multi-turn conversational agents with context windows up to 128k tokens. Meridian handles session hydration, tool calling, and streaming responses out of the box.

Recommended model

Claude 3.5 Sonnet or GPT-4o

Lowest latency for real-time chat, strong instruction following, and native tool-use support.

View example

RAG

Retrieval-augmented generation over your own documents. Ingest PDFs, markdown, or codebases, chunk with a sliding window, embed with your provider of choice, and query via hybrid search.

Recommended model

Claude 3.5 Sonnet + text-embedding-3-large

Sonnet handles long retrieved contexts reliably; OpenAI embeddings give the best recall on technical content.

View example

Code Copilot

Inline completions, refactoring, and PR review. Meridian streams edits via the diff protocol so your IDE can apply changes without round-trips.

Recommended model

Claude 3.5 Sonnet or GPT-4o

Top-tier HumanEval scores, consistent code generation, and fast streaming for real-time completions.

View example

Batch Summarization

Process thousands of documents, transcripts, or support tickets in parallel. Meridian fans out across your rate limits and merges results into structured JSON.

Recommended model

GPT-4o mini or Claude 3.5 Haiku

Cost-effective at scale; 4o mini is 90% cheaper than 4o with near-identical summarization quality.

View example

Vision OCR

Extract text, tables, and handwriting from images or scanned PDFs. Meridian preprocesses with sharpening and deskew before sending to the model.

Recommended model

GPT-4o or Claude 3.5 Sonnet

Both models read complex layouts, handwritten notes, and low-contrast scans with >99% accuracy.

View example

Voice Agent

Real-time speech-to-speech pipelines. Meridian bridges Deepgram or Whisper for STT, your LLM for reasoning, and ElevenLabs or Play.ht for TTS — all over a single WebSocket.

Recommended model

GPT-4o (audio preview) or Claude 3.5 Sonnet + Deepgram

GPT-4o audio natively handles tone and pacing; the Sonnet + Deepgram stack gives you provider flexibility.

View example