AI Recommendation Engine API in India

AI-powered recommendations, built on your own catalog

Generic "customers also bought" widgets leave money on the table. unoblox lets you build a recommendation layer that reasons about a user's actual stated interests and history — not just co-purchase statistics — through one OpenAI-compatible endpoint, billed in rupees.

Where LLM recommendations beat collaborative filtering

Classic collaborative filtering needs thousands of interactions per item before it recommends anything useful, and it has little to say about a brand-new user or a brand-new product (the "cold-start" problem). A language model can reason from a short profile plus a candidate list on day one, and it can explain why it picked an item — useful for both user trust and debugging.

Best models for recommendations

Model	Input ₹/1M	Output ₹/1M	Fit
Qwen3 235B-A22B	₹9.07	₹55.44	High QPS, low cost per call
GPT-4.1	₹201.6	₹806.4	Strong reasoning over long candidate lists
Claude Sonnet	₹201.6	₹1008	Nuanced taste and preference modelling

Prices per 1M tokens — check /models for live rates.

A minimal recommender

from openai import OpenAI

client = OpenAI(
    api_key="ub-gw-...",
    base_url="https://api.unoblox.ai/v1"
)

user_profile = {
    "name": "Priya",
    "interests": ["sci-fi", "psychology", "travel"],
    "past_purchases": ["Dune", "Thinking, Fast and Slow"],
    "budget": "under ₹500"
}

candidates = [
    "Foundation by Asimov",
    "Sapiens by Harari",
    "Atomic Habits by Clear",
    "The Midnight Library"
]

response = client.chat.completions.create(
    model="openai/gpt-4-1",
    messages=[{
        "role": "user",
        "content": "Given this profile: " + str(user_profile) + " and these candidates: " + str(candidates) + ", recommend the top 2 and explain why, in one line each."
    }]
)

print(response.choices[0].message.content)

Designing the candidate set

The model is only as good as what you hand it — send too many candidates and cost and latency climb; send too few and recommendations feel repetitive.

1. Pre-filter with rules: stock availability, budget cap, category match (keep the candidate list under ~50 items).
2. Let the model rank, not search: pass the shortlist and ask it to order and explain, rather than searching your whole catalog per call.
3. Blend in trending items: mix in one or two popular picks so new users without history still see something relevant.
4. Cache by user segment: group similar users to cut repeat-call cost for near-identical profiles.

Where teams use this

• E-commerce product recommendations and "you might also like" rails
• Streaming content curation (shows, music, podcasts)
• Job boards, matching postings to a candidate's stated preferences
• Learning platforms, sequencing the next course or module
• Marketplace service discovery (matching buyers to sellers)

What it costs

A single recommendation call — profile plus a 20–30 item candidate list — typically runs 300–600 tokens. At GPT-4.1 rates that's roughly ₹0.05–₹0.10 per call; on Qwen3 235B-A22B it drops to a fraction of a paisa, which matters once you're serving recommendations on every page load rather than once per session.

Frequently asked questions

Q: How do I avoid recommending the same items to everyone? A: Feed in the user's actual history and stated preferences, not just category. Ask the model to weight novelty, and rotate in less-popular candidates periodically.

Q: What do I do for a brand-new user with no history? A: Use onboarding questions (a few preference prompts) to bootstrap a profile, and blend in trending or editor's-pick items until real behavioral data accumulates.

Q: How should I test whether recommendations are actually working? A: Run an A/B test — one arm sees AI recommendations, a control arm sees your existing logic. Compare click-through and conversion rate, not just how "relevant" the picks look.

Q: Can recommendations be in Hindi or a regional language? A: Yes. Claude Sonnet and GPT-4.1 both handle Indian languages well — write item descriptions and prompts in the language your users browse in.

Q: What if the model recommends something out of stock? A: Filter the candidate list before the call, not after — never hand the model items it shouldn't be able to recommend.

Q: Can I combine this with existing collaborative filtering? A: Yes, and it's often the best setup — use collaborative filtering to build the candidate shortlist, then let the model do the final personalized ranking and explanation.

Get started in rupees → https://unoblox.ai/sign-in