Relia

Financial relevance scoring — is this news article actually about this company?

Relia is a hosted API. You send an entity (e.g. NVDA) and a batch of news articles and get back a calibrated relevance score in [0, 1] for each one — telling apart an article that moves the company from one that merely name-drops it. It reproduces the judgment of a panel of frontier LLM raters, but runs as a tiny purpose-built model: ~100× faster and at a fraction of the per-token cost.

Current model: relia-1 — a LoRA-adapted Qwen3-Embedding-0.6B bi-encoder with a cumulative-link ordinal head.

---

Why Relia

	Relia-1	Frontier LLM raters
Accuracy	86.5% exact / 100% within-1-bin agreement with a 3-LLM consensus	88–90.5% (the ceiling)
Speed	~18 ms/article batched (measured)	~1.8 s per article, per API call
Cost	$0.0025 / call (up to 50 articles/call, first 500 free), no per-token metering	per-token billing, every call
Dependencies	one call to one model	an API round-trip (or three, for a consensus)

The striking part: Relia agrees with the LLM consensus (86.5%) more than the LLMs agree with each other (78–80.5%). It sits inside the raters' own noise band.

Charts are generated from aggregate metrics by validation/make_charts.py. Full method and numbers in validation/README.md.

---

What you get

For each article Relia returns:

• score — expected relevance in [0, 1]
• level — the argmax relevance bin (0 irrelevant · 1 low · 2 medium · 3 high)
• probs — the full distribution over the four bins

Example

Request

curl -X POST https://<your-deployment>/v1/score \
  -H "Content-Type: application/json" \
  -d '{
    "entity": "NVDA",
    "articles": [
      {"id": "a1", "title": "Nvidia unveils new Blackwell GPU for AI data centers",
       "body": "Nvidia announced its next-generation Blackwell architecture, boosting AI training performance ..."},
      {"id": "a2", "title": "Local bakery wins county pie contest",
       "body": "A small-town bakery took home first prize at the annual county fair ..."}
    ]
  }'

Response

{
  "model": "relia-1",
  "entity": "NVDA",
  "scores": [
    {"id": "a1", "score": 0.875, "level": 3, "probs": [0.0, 0.0, 0.001, 0.999]},
    {"id": "a2", "score": 0.125, "level": 0, "probs": [1.0, 0.0, 0.0, 0.0]}
  ]
}

The Blackwell story lands in the top bin (level 3); the bakery story in the bottom bin (level 0) — for the same entity, purely from the text.

You supply an entity symbol; Relia uses its own curated profile for that entity — you never author or upload profiles. List supported entities with GET /v1/entities; an uncovered symbol returns 404 (no silent fallback). See docs/COVERAGE.md.

Full API reference: docs/API.md.

---

Pricing & rate limits

• $0.0025 per call — 1 call = 1 entity + up to 50 articles.
• First 500 requests free, no time limit.
• 120 requests / minute per API key.

Get your API key. Full pricing and limits in docs/PRICING.md.

---

About this repository

This repo is the public interface for the Relia API — the request/response schema, API reference, coverage, benchmarks, and pricing. Relia is a fully hosted service; there is nothing to install or deploy. Get an API key and call the endpoint.

Licensing

This repository contains the public documentation and examples for the Relia API. Copyright (c) 2026 Aleta / AletaIndex. All rights reserved. The relia-1 model — its weights and curated entity profiles — is proprietary and is not distributed here; access is provided via the hosted Relia API under its Terms of Service.