llm-core

Appearance

Unified Media Inputs ​

One of the hardest parts of working with LLMs is juggling input formats. Some APIs want base64 strings, others want Data URLs, and others want raw buffers.

llm-core handles this for you. You pass what you have; the framework makes it work.

The Principle: "Just Pass It" ​

The library uses a Universal Content Normalizer that runs before any adapter sees your data. This means you can be agnostic about how your data is stored.

Images ​

Whether you have a remote URL, a local file path, or a raw Buffer, the syntax is the same.

ts
// 1. Remote URL
await workflow.run({
  image: "https://example.com/chart.png",
});

// 2. Data URI
await workflow.run({
  image: "data:image/png;base64,iVBORw0KGgo...",
});

// 3. Raw Buffer (Node.js) / Uint8Array
const myBuffer = fs.readFileSync("chart.png");
await workflow.run({
  image: myBuffer, // Checksum & Base64 handling is automatic
});

Mixed Content (Multi-modal) ​

You don't need to manually construct complex API objects. Use the toMessageContent helper to unify text and images.

ts
import { toMessageContent } from "@geekist/llm-core/adapters";

const input = toMessageContent([
  "Here is the design:",
  { type: "image", url: "https://..." },
  "What do you think?",
]);

// Result is standardized for ALL providers (OpenAI, Anthropic, etc.)
await workflow.run(input);

Advanced: Buffers & Binary Data ​

When you pass a Buffer or Uint8Array, the framework:

  1. Auto-Detects Mime Type (where possible) or accepts an explicit one.
  2. Converts to Base64 lazily, only when the specific provider adapter needs it.
  3. Prevents Copying: Large buffers are handled by reference until the last possible moment.
ts
// Explicit MIME type if needed
await workflow.run({
  type: "image",
  data: myBuffer,
  mimeType: "image/png",
});

Next Steps ​

Unified Media is just one part of the "Universal Normalisation" story.