About this site
Parakhi — from parakh, to assay or test the purity of a coin — is an open-source civic-tech project that breaks down everyday Indian consumer products: how much of the price is raw materials, how much is tax, how much is the retailer's margin, and how much went to a country other than India. Kya hai andar?
Most "AI gives you a number" sites lie with confidence. We don't. Every number on this site carries a confidence indicator and a source tier. If we don't know something, we say so.
How we estimate
- Identify the product — barcode via Open Food Facts, URL via Open Graph, or free text via a small language-model normalizer.
- Classify it against one of our curated category templates (e.g. "packaged biscuits"). If no template fits, we show "category not yet supported" — we do not invent breakdowns.
- Compute the breakdown deterministically from the category template — no language model touches the numbers. Same template, same product, same answer. Reproducible.
- Cache the result in our DB so subsequent queries are free.
Division of labor
We use language models only for the parts where they're robust — text normalization and classification. The numbers themselves come from category templates plus arithmetic. Specifically:
- Template authors set the raw-material composition, origin probabilities, and typical cost bands for a category, citing public sources.
- A live data layer keeps the inputs fresh: GST rates from the CBIC HSN schedule, brand-to-parent-country from Wikidata, commodity prices from data.gov.in. No human re-types these.
- Code computes the per-product breakdown — labels, ranges, rupee amounts, both scores below, and source tiers.
- A small LLM (only on our hosted site) resolves typed queries and picks the matching category. It never sees or sets a number. Self-hosted forks can run fully LLM-free.
Two scores, two questions
"Made in India" smuggles three different questions into one number. We split them.
🇮🇳 Indian Value Capture (the headline)
"Where does my rupee go?" The MRP-weighted share that flows to Indian sources across every cost bucket — raw materials, packaging, manufacturing, logistics, retailer + distributor margin, brand margin, advertising, and brand profit. Each bucket is multiplied by its probability of Indian origin.
GST is deliberately excluded. Tax going to the government is a different question from value reaching producers, workers, and retailers — and since GST ranges from 0% (milk) to 40% (aerated drinks), including it would distort every cross-category comparison. We show GST as its own number instead.
IVC is sensitive to things people miss: aluminum cans imported from the Middle East, royalty flowing to a foreign parent, specialty enzymes from China. When a brand has a known foreign parent (via Wikidata), its brand-profit share is attributed abroad automatically.
Composition Made-in-India (the secondary chip)
"What is it physically made of, from where?" Weighted probability of Indian origin across raw materials only, by composition share. A bottled water or a Diet Coke scores high here — it's mostly Indian water by volume — even when its brand value flows abroad. Useful as a sanity check, never the headline.
🏛️ GST (shown separately)
The rupees per pack that go to the government as tax, sourced from the CBIC HSN→rate schedule. Tier 1, always.
Where we can't justify a foreign share from a public source, we default to fully Indian — we'd rather under-claim foreignness than invent it.
Source tiers
- Tier 1 — Hard data (HSN/GST schedules, government commodity prices, RoC filings).
- Tier 2 — Structured open data (Open Food Facts, public industry reports).
- Tier 3 — Public web sources (brand websites, news, annual reports).
- Tier 4 — LLM reasoning, always shown with ranges.
What we don't do
- No ads.
- No login, no accounts, no tracking beyond rate-limit hashes.
- No invented breakdowns for products without a template.
- No point estimates where ranges are honest.
Open source
Code, templates, and prompts live on GitHub. If you spot an error, send a correction — we read every one.