ObjectiveAI is a platform for Functions - remote scoring/ranking pipelines powered by ensembles of LLMs. Define a Function, train it on your data, and call it from anywhere.

Website | API | Discord | npm | crates.io

• Ranking search results, candidates, or recommendations
• Scoring content quality, code reviews, or support tickets
• Classifying documents or user intents
• Simulating how a specific person or group would decide

npm install objectiveai

[dependencies]
objectiveai = "0.1.1"

An Ensemble is a group of LLMs that vote together. Each LLM can have its own personality prompt, temperature, output mode, etc.

{
  "llms": [
    {
      "model": "openai/gpt-4o",
      "output_mode": "json_schema",
      "prefix_messages": [
        { "role": "system", "content": "You are a rational skeptic. You select responses which are grounded in logic." }
      ],
      "count": 2
    },
    {
      "model": "anthropic/claude-sonnet-4-20250514",
      "output_mode": "tool_call",
      "suffix_messages": [
        { "role": "system", "content": "You are an emotional believer. You select responses that make you feel good." }
      ],
      "count": 1
    }
  ]
}

Give the ensemble a prompt and possible responses. Each LLM votes for what it thinks is the best response, and votes are combined with weights to produce scores.

Prompt: "What color is the sky?"
Responses: ["blue", "green", "red", "yellow"]

-> Scores: [0.85, 0.05, 0.05, 0.05]  (sums to 1)

LLMs are inherently probabilistic - the sampler makes the final discrete choice. ObjectiveAI bypasses the sampler entirely using logprobs to capture the model's full preference distribution.

Instead of asking "which is best?" and getting one answer, we extract the probability the model assigns to each option simultaneously. If the model is 70% confident in A and 30% in B, we capture that nuance rather than losing it to sampling.

Traditional: Model outputs "A" (loses the 30% signal for B)
ObjectiveAI: Model vote = [0.70, 0.30, 0.00, 0.00] (full distribution)

For large response sets exceeding logprobs limits, a prefix tree structure captures preferences in stages - the tree width matches logprobs count (typically 20), enabling voting over hundreds of options while preserving probability information at each level.

Functions are the main interface. They're composable scoring pipelines hosted on GitHub as function.json files.

Input -> [Tasks] -> Score

Reference by owner/repo:

objectiveai/sentiment-scorer

ObjectiveAI doesn't fine-tune models - it learns weights over your ensemble.

Give it a dataset of inputs and expected outputs. It optimizes the weights to match, producing a Profile you can reuse. Profiles are GitHub-hosted as profile.json.

objectiveai/
├── objectiveai-rs/           # Rust SDK (core crate)
├── objectiveai-api/          # API server (run locally or import as library)
├── objectiveai-rs-wasm-js/   # WASM bindings
├── objectiveai-js/           # TypeScript SDK
└── objectiveai-web/          # Web interface

MIT