LLM & model discovery

TypeScript SDK methods for the governed LLM surface — chat, responses, embeddings, rerank, tokenization, model artifacts — plus model catalog and runtime discovery.

These AlgentaClient methods cover the governed LLM surface (chat completions, responses, embeddings, rerank, tokenization, and model-artifact resolution) plus the discovery calls that tell you which models and which engine build the instance actually serves. Method names are camelCase; bodies stay snake_case. Every model call is routed per the instance catalog — a model not in listModels() cannot be called.

Chat and responses

chatCompletions runs a message list to completion; streamChatCompletions returns an AsyncGenerator of chunk responses. The responses and streamResponses pair works the same way over the responses API. model is optional — omit it to use the instance default.

import { AlgentaClient } from "algenta-sdk";

const client = new AlgentaClient({ apiKey: process.env.ALGENTA_API_KEY });

// Single completion.
const completion = await client.chatCompletions({
  messages: [
    { role: "system", content: "You are terse." },
    { role: "user", content: "Summarize Q3 in one line." },
  ],
});
console.log(completion.choices[0].message.content);

// Streamed completion.
for await (const chunk of client.streamChatCompletions({
  messages: [{ role: "user", content: "List three risks." }],
})) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}

Method	HTTP	Notes
chatCompletions(request)	POST /v1/chat/completions	messages with role/content; returns a ChatCompletionsResponse. Sends stream: false.
streamChatCompletions(request)	POST /v1/chat/completions	AsyncGenerator of ChatCompletionsStreamChunkResponse. Sends stream: true.
responses(request)	POST /v1/responses	input string or array; returns a ResponsesResponse. Sends stream: false.
streamResponses(request)	POST /v1/responses	AsyncGenerator of ResponseStreamEventResponse. Sends stream: true.

Embeddings and rerank

embeddings returns one vector per input; embeddingSimilarity scores two vectors directly; rerank orders documents (each with a precomputed embedding) against a query_embedding.

// Embed text.
const embedded = await client.embeddings({
  input: ["empty cart at checkout", "payment declined"],
});
console.log(embedded.data.length, embedded.dimensions);

// Rerank documents against a query embedding.
const reranked = await client.rerank({
  query_embedding: embedded.data[0].embedding,
  documents: [
    { id: "doc_1", embedding: embedded.data[0].embedding, text: "empty cart at checkout" },
    { id: "doc_2", embedding: embedded.data[1].embedding, text: "payment declined" },
  ],
  top_n: 2,
});
console.log(reranked.data.map((d) => [d.id, d.rank, d.score]));

Method	HTTP	Notes
embeddings(request)	POST /v1/embeddings	input string or array; optional dimensions.
embeddingSimilarity(request)	POST /v1/embeddings/similarity	left and right vectors; returns a score.
rerank(request)	POST /v1/rerank	query_embedding plus documents (id, embedding, text?); optional top_n.

Tokenization

Method	HTTP	Notes
tokenize(request)	POST /v1/tokenize	Returns the tokens array and token_count for the input.
countTokens(request)	POST /v1/count_tokens	Returns just the token_count for the input.

Model artifacts

Method	HTTP	Notes
resolveArtifactBridge(request)	POST /v1/artifacts/resolve	Resolves a model artifact by repo_id and filename; defaults to local_files_only: true.

Model catalog and runtime discovery

Before you route to a model, ask the instance what it can run. listModels returns the executable catalog; getContract returns the machine-readable platform contract (and falls back to /openapi.json if the contract endpoint is absent); the runtime getters read the signed manifest and validation artifacts for the active build.

const models = await client.listModels();          // GET /v1/models
const contract = await client.getContract();        // GET /v1/meta/contract
const manifest = await client.getRuntimeManifest(); // GET /v1/runtime/manifest
console.log(models.data.length, manifest);

Method	HTTP	Notes
listModels()	GET /v1/models	The catalog of models this instance can execute.
getContract()	GET /v1/meta/contract	Machine-readable platform contract; falls back to /openapi.json.
getRuntimeManifest()	GET /v1/runtime/manifest	Signed manifest of the active runtime build.
getRuntimeModules()	GET /v1/admin/runtime/modules	Registered runtime-library modules.
getRuntimeBenchmarks()	GET /v1/admin/runtime/benchmarks	Published runtime benchmarks.
getRuntimeReleaseValidation()	GET /v1/admin/runtime/validation	Release-validation proof for the build.

getContract, getRuntimeManifest, and the runtime getters validate their payloads against the runtime-proof contract. An invalid or tampered response raises an error rather than returning unverified data.

Next steps

Connect data sources

Create connectors, onboard datasets, and triage repositories.

Simulate and query

Run Monte Carlo decisions and governed data queries.

Discover the API

Bootstrap from the health, contract, and models endpoints.