MetrIQ

INDIAN STANDARDS · BIS SERVICES · EVIDENCE-BACKED

Which Indian Standard governs this product?

Ask in plain English, Hindi or Telugu — or photograph the label.
MetrIQ answers only from verified BIS records, shows why each result matched,
and links the official BIS page behind every answer.

_{▶ Click to play the 40-second demo — every scene is the real app doing its job.}

505 VERIFIED INDIAN STANDARDS

98.1% RIGHT STANDARD IN TOP 5

0 INVENTED STANDARD NUMBERS

3 LANGUAGES · EN · हिन्दी · తెలుగు

_{Recall@5 over 149 queries derived from BIS's own product ↔ standard pairings (backend/tests/data/eval_baseline.json). "0 invented" = zero standard numbers outside the knowledge base across every adversarial and prompt-injection query.}

---

The one rule

Retrieved BIS information is the source of truth — not the language model.

natural language  →  query understanding  →  BIS knowledge retrieval
                  →  evidence ranking     →  grounded explanation
                  →  answer + evidence + official source + next steps

Every result is produced by deterministic code. A language model only explains what retrieval already found — it never picks a standard. MetrIQ never invents an Indian Standard number, a clause, a fee, a test or a compliance outcome, and it gives no PASS/FAIL verdict anywhere. When the evidence is not strong enough, it says so and abstains.

---

A tour of the app

01 · PRODUCT → STANDARD Describe a product, get its Indian Standard "LED bulb" → IS 16102 (Part 1) at HIGH confidence, with the matched words and a plain "Why this result?" built from the retrieval signals themselves.	02 · HONEST ABSTENTION No evidence, no answer "shampoo" is not on BIS's compulsory-certification lists, so MetrIQ says what its data covers — and that this does not mean no Indian Standard exists.
03 · STANDARD PASSPORT One canonical page per standard Identity, edition, currency, legal status, clauses and the certification route — each line quoting its BIS source. Prints to PDF.	04 · LABEL INSPECTION Photograph the label instead Fully local OCR reads the package; every declared value links back to the exact box it came from, with its confidence and extraction method.
05 · TESTING LABORATORIES Labs BIS itself lists for the standard From a dated snapshot of BIS LIMS. Alphabetical, never ranked; validity is stated "at snapshot", never as current accreditation.	06 · HALLMARKING / HUID Observed, never authenticated Reads a potential HUID and purity mark from a photo and points to the BIS Care App for real verification. No "verified" state exists in the code.

_{07 · ASK IN YOUR LANGUAGE}
Hindi and Telugu in, the same verified evidence out
_{The knowledge base is never translated — standard numbers, schemes and URLs stay exactly as BIS publishes them.}

---

How it works

flowchart LR
    Q["Question<br/>EN · HI · TE"] --> L[Language layer<br/>deterministic]
    P["Label photo"] --> O[Local OCR<br/>PP-OCRv3 · ONNX]
    O --> D[Declarations<br/>regex + keywords]
    L --> R[Retrieval engine<br/>weighted lexical · MatchReasons]
    D --> R
    R --> S[Verified BIS records<br/>505 standards · QCOs · certification · labs]
    S --> W[Why this result?<br/>no model]
    S --> E[Grounded explanation<br/>OpenRouter · guarded]
    W --> A[Answer + evidence<br/>+ official source]
    E --> A

    style S fill:#eceffe,stroke:#2246ef,color:#17181b
    style E stroke-dasharray: 4 4

Loading

Layer	What it does	Model?
Retrieval (app/retrieval/)	Transparent weighted scoring over title / keywords / standard number / content; every point recorded as a MatchReason	no
Why this result? (app/product.py)	Built from retrieval's own reasons — never generated	no
Certification journey (app/certification_journey.py)	Every step is a word-for-word quote from a verified record, with its URL	no
Standard Passport (/standard/:id)	Composes identity, currency, QCO status, clauses, route, labs	no
Grounded answer (app/rag.py)	Explains the retrieved records; a guard withholds any invented IS number, HUID, URL, fee or verdict	yes — replaceable

---

Quickstart

# 1 · Backend  (Python 3.11+)
cd backend
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
alembic upgrade head                      # needs PostgreSQL, DATABASE_URL default postgresql+psycopg:///metriq
uvicorn app.main:app --port 8000          # http://127.0.0.1:8000/docs

# 2 · Frontend
cd frontend
npm install
npm run dev                               # http://localhost:5173  (proxies /api → :8000)

Grounded answers need an OpenRouter key in backend/.env (gitignored, server-side only):

OPENROUTER_API_KEY=sk-or-v1-...
OPENROUTER_MODEL=inclusionai/ling-3.0-flash-vl:free           # copilot
OPENROUTER_GROUNDED_MODEL=inclusionai/ling-3.0-flash-vl:free  # /ask + certification

Without a key, everything deterministic still works — retrieval, Why this result, the certification journey, the Passport, labs and OCR — and /ask falls back to evidence-only text.

PostgreSQL setup (Arch Linux)

sudo pacman -S --needed postgresql
sudo -iu postgres initdb -D /var/lib/postgres/data
sudo systemctl enable --now postgresql
sudo -iu postgres createuser -s "$USER"
createdb metriq && createdb metriq_test          # app database + test database
cd backend && ./.venv/bin/alembic upgrade head

Only saved inspections need the database; the /inspections endpoints return 503 until it is available.

Optional: local model (LM Studio) and vision key

LM Studio is used in exactly two places: the inspection pipeline's product-identification fallback and Laboratory search's explain=true path. Load a small instruction model (default qwen/qwen3-4b) and start its server on port 1234:

export LM_STUDIO_BASE_URL=http://127.0.0.1:1234/v1
export LM_STUDIO_MODEL=qwen/qwen3-4b

Visual product understanding uses its own key, so its quota and outages are independent:

OPENROUTER_VISION_API_KEY=sk-or-v1-...     # NOT the same as OPENROUTER_API_KEY
VISION_MODEL=inclusionai/ling-3.0-flash-vl:free

Free-tier guards (OPENROUTER_DAILY_LIMIT 45, OPENROUTER_MINUTE_LIMIT 15, VISION_MAX_IMAGES, …) refuse a request locally before it reaches the network. A request is sent only when you press a question — never on page load.

---

Tests & checks

cd backend
./.venv/bin/python -m pytest -q                  # 395 tests (needs the metriq_test database)
./.venv/bin/python scripts/check_knowledge.py    # knowledge-base validation + coverage
./.venv/bin/python scripts/eval_retrieval.py     # offline retrieval eval vs committed baseline
./.venv/bin/python scripts/verify_snapshot.py    # source drift report (never writes the KB)

cd ../frontend
npx tsc --noEmit && npm run build

Model-dependent tests use stubs — no live model call is made and no quota is spent.

Retrieval baseline
recall@1	85.6%
recall@5	98.1%
correct abstention	24.4%
false match	13.4%

---

What's in the knowledge base

505	verified Indian Standards, transcribed from BIS's "Products under Compulsory Certification" (Scheme I / II) with source URL and verification date
1,537	OCR'd clause records from 31 standards — citation-only, labelled as unverified OCR text, never used for retrieval
29	Quality Control Order rows — status comes from which BIS table a row is in, never from today's date
1,205	BIS LIMS laboratory records (245 labs · 83 cities) — 86 of 505 standards have a listed lab
70	certification, FAQ, hallmarking, Legal Metrology, consumer and general BIS records

This is a focused, curated dataset — not complete BIS coverage. About 273 of the ~769 products notified under compulsory certification sit in Quality Control Orders outside the two BIS listing pages, and MetrIQ does not guess them.

---

Deep dive

API endpoints

Endpoint	What it does	Model
GET /health	liveness	—
GET/POST /search	deterministic lexical retrieval over the BIS knowledge base	no
POST /product-standard	Product → candidate Indian Standard + "Why this result?"	no
POST /ask	grounded BIS Q&A in English / Hindi / Telugu (also serves Hallmarking)	OpenRouter
POST /certification-guidance	certification journey + optional grounded explanation (explain=false skips it)	OpenRouter, optional
POST /laboratory-search	laboratories for a standard or product	LM Studio, optional
GET /standard-passport/{record_id} · /standard-passport/lookup	the Standard Passport for one standard (a number without a year lists every held edition)	no
POST /product-context	one server-derived context across all features	no
POST /evidence-graph	read-only graph of relationships already established	no
POST /inspection/ocr · /inspection/analyze	label photo(s) → OCR → declarations → product → standard	LM Studio fallback only
POST /inspections · GET /inspections[/{id}]	save / list / read immutable inspection records	no
GET /inspections/{id}/report.pdf	evidence-backed PDF built from the stored record	no
POST /copilot/explain	grounded explanation of a finished result	OpenRouter

Label inspection pipeline

IMAGE → OCR REGION → DECLARATION → PRODUCT CLUE → PRODUCT     → STANDARD CANDIDATE → BIS SOURCE
       (id, bbox,    (field, value, (text + its   (KB product   (IS 16102 (Part 1),   (verified
        confidence)   status)        OCR regions)  description)  why this result)      record)

One package is a single photo (image, optional side) or several photos of the same package (images + sides = FRONT / BACK / LEFT / RIGHT / TOP / BOTTOM / UNKNOWN). Each photo is OCR'd separately and every value keeps the photo and region it came from; a photo that cannot be read is reported as failed, and unphotographed sides as not uploaded — never as missing.

Products and standards come only from data/knowledge/ through the same retrieval engine as Product → Standard. A product match needs a product phrase of that record on the label; an IS number printed on the label is one more signal, never proof on its own. Anything weaker is REVIEW, with the reason. Declarations are DETECTED / UNCERTAIN / NOT_DETECTED in the uploaded photos — never "legally missing".

app/escalation.py then states, deterministically, every point of the evidence chain the photos could not establish (unreadable photo, product not identified, several candidates, no verified standard, HUID, conflicting sides, …). It changes no result and produces no verdict.

Saved inspections are stored in PostgreSQL with their photos and are immutable — a database trigger rejects any update. History, a dashboard and a PDF report all read the stored record; nothing is recomputed.

curl -s -F "image=@samples/ocr-labels/synth_led-lamp.png" \
  http://127.0.0.1:8000/inspection/analyze | jq '.product, [.standards[].standard_number]'

Sample (samples/ocr-labels/)	Identified as	Standard candidate
synth_clean-declaration.png	Roasted Bengal Gram	IS 18140:2023
synth_led-lamp.png	Self-ballasted LED lamps	IS 16102 (Part 1)
synth_electric-kettle.png	Electric Kettles and Jugs	IS 367:1993
real_* (Wikimedia Commons)	real-world labels, incl. hard cases	OCR stress tests

Hallmarking & HUID — observed, never authenticated

MetrIQ can identify and explain observable hallmark/HUID evidence, but it does not authenticate a physical jewellery item's hallmark, HUID, jeweller registration or AHC status.

• verification_status is only NOT_VERIFIED or NOT_DETECTED; there is no AUTHENTIC, VERIFIED or CERTIFIED state anywhere in the code.
• The three BIS marks — BIS logo, purity/fineness, HUID — are each DETECTED / NOT_DETECTED / UNCERTAIN / NOT_SUPPORTED. The BIS logo is always NOT_SUPPORTED: it is a graphic and OCR reads text.
• "Not detected" is about the photograph, never a finding that the article lacks the mark.
• Printed text like "HUID VERIFIED" is recorded as an untrusted claim and changes nothing.
• A HUID the user types is compared with the OCR text as a string only.
• Real verification is pointed to the BIS Care App, quoted from verified BIS records.

Testing laboratories

MetrIQ identifies laboratories from verified laboratory evidence. It does not independently establish a laboratory's current accreditation, scope, availability or operational status.

Search by standard, by a question naming a standard, or by product (which reuses Product → Standard: "Where can I test an electric kettle?" → IS 367:1993 → the labs BIS LIMS lists for it). A lab is relevant only because BIS lists it against that standard. Each result carries a deterministic why (STANDARD_LISTED, PRODUCT_LISTED, NAME_MATCH, CITY_MATCH), results are alphabetical, and validity is VALID_AT_SNAPSHOT / EXPIRED_AT_SNAPSHOT / NOT_STATED. A different edition is a different standard, so IS 14543 (2016) and (2024) labs are never merged. Addresses, phones, emails and NABL status are not held and never supplied.

The snapshot (data/laboratories.json, retrieved 2026-09-19) is built by backend/scripts/fetch_lims_laboratories.py; the app never calls LIMS at runtime.

Multilingual — English · Hindi · Telugu

The language of interaction changes, the source of truth does not.

query (any language) → script detection (no model) → known terms rewritten to canonical English
                     → the SAME retrieval engine → the SAME verified records
                     → grounded answer written in the user's language

/ask, /certification-guidance and /laboratory-search take language = auto / en / hi / te. The alias table (app/language.py) only maps spellings of products and BIS terms that exist in the knowledge base. Standard numbers, scheme names, document names and URLs are reproduced exactly, never translated. If retrieval finds nothing, the assistant abstains in the user's language. Hinglish / Tanglish work.

Certification journey

POST /certification-guidance answers "what certification process do I follow?". Nothing in the journey is written by MetrIQ — each step is a word-for-word quote from a verified record with its official URL. The scheme is read two independent ways (the BIS listing the standard was transcribed from, and any certification record naming it); if they disagree, the conflict is shown and MetrIQ picks neither.

Status	Meaning
VERIFIED	one standard, a route established, every documented step present
PARTIAL	several candidates, a conflict, missing steps, or hallmarking
INSUFFICIENT	no verified record states a route

It is guidance about the route for a product type — never a statement that a product or licence is certified. Fees, processing times and required documents are never stated; they live only in the BIS documents it links to.

Copilot, product context & evidence graph

Copilot — explains evidence the deterministic system produced; it never establishes standards, compliance, lab status, certification applicability or HUID authenticity. After the model replies, a deterministic guard replaces the text with MetrIQ's own sentence if it cites an IS number, HUID or URL not in the context, claims authentication, states any PASS/FAIL or "compliant" verdict, ranks or vouches for a laboratory, or states a fee not in the evidence. The key stays on the server — there is no VITE_ variable for it.

Product context (app/product_context.py) — connects what each feature already established into one view, with each feature AVAILABLE / NOT_AVAILABLE / NOT_APPLICABLE / UNCERTAIN. It is a composition: no classifier, no ranking, no model. Conflicts are shown, never resolved.

Evidence graph (app/evidence_graph.py) — a read-only projection of relationships already established (OCR region → declaration → product → standard → requirement / certification / laboratory / source). 11 node types, 8 edge types, each edge carrying an explanation taken from the evidence. No model, no retrieval, no verdict vocabulary; nothing downstream reads it.

Retrieval scoring reference

Each query term is matched against a record's fields and the weights are summed (RetrievalConfig in app/retrieval/engine.py): standard-number match 8 (12 if the year also matches), title 4, keyword 3, category hint 2, document name 1.5, reference 1, content mention 1.

Confidence of the top hit: high ≥ 7.5, medium ≥ 4.0, low ≥ 1.0, else none. If the top hit covers less than ~⅓ of the query terms, confidence is capped at low. When nothing matches: confidence: "none", abstained: true, empty results.

Inspection product identification keeps a result only when the label contains a multi-word product phrase of the record, so a single generic word ("water", "gram") can never pull in a standard.

Repository layout

backend/            Python · FastAPI
  app/              retrieval/, rag.py, product.py, certification*.py, laboratory.py, lab_registry.py,
                    language.py, clauses.py, clause_groups.py, standard_currency.py, boundary.py,
                    ocr.py, declarations.py, pipeline.py, hallmark.py, copilot.py, openrouter.py, …
  migrations/       Alembic (PostgreSQL)
  scripts/          check_knowledge, eval_retrieval, verify_snapshot, build-time fetch_* tools
  tests/            plain-Python runners, bridged to pytest
data/knowledge/     the knowledge base — one JSON file per category
data/laboratories.json   BIS LIMS snapshot
samples/ocr-labels/ sample label images
frontend/           React 19 · TypeScript · Vite · Tailwind v4
docs/history/       phase-by-phase design decisions

---

_{BUILT FOR SMART INDIA HACKATHON · PROBLEM STATEMENT SIH26107
MetrIQ is a research prototype. It is not affiliated with or endorsed by the Bureau of Indian Standards, and it does not make legal or compliance determinations.
Always confirm against the official BIS source it links.}