Step 0: Full Project Analysis

0.1 What the Project Is

Project Name & Purpose

21-BIS-Compliance-Engine is an air-gapped, offline-first Regulatory Compliance and Gap Analysis Platform designed to automate statutory product compliance audits under two major Indian regulatory regimes:

1. Bureau of Indian Standards (BIS) Scheme-I (ISI Mark Certification) governed by the Bureau of Indian Standards Act, 2016 and statutory Quality Control Orders (QCOs).
2. Automotive Industry Standards (AIS) Homologation governed by Central Motor Vehicles Rules (CMVR), 1989 (Rule 126), covering automotive safety components, electric vehicle powertrains, and location-tracking systems (ARAI/ICAT/CIRT certification).

Problem Solved

Manufacturing enterprises, MSMEs, importers, and compliance officers face significant regulatory hurdles when bringing industrial or automotive hardware to the Indian market. Determining whether a specific manufactured item falls under a mandatory Quality Control Order (QCO), identifying the corresponding standard number (e.g., IS 12269 or AIS 156), establishing the designated testing laboratory, and determining factory audit readiness currently requires weeks of manual legal review. Non-compliance results in port confiscations, factory shutdown notices, or heavy statutory fines.

This platform automates this entire audit lifecycle in sub-50 milliseconds by combining:

• A fine-grained Scikit-Learn product classification model.
• An ONNX-accelerated dense semantic vector retrieval engine (FastEmbed BAAI/bge-small-en-v1.5 over a FAISS IndexFlatIP vector index).
• A domain-specific multi-tiered regulatory guardrail.
• An automated weighted gap analysis engine computing statutory readiness scores against Scheme-I and AIS-037 factory compliance criteria.
• One-click statutory CSV audit report export.

Target Users & Personas

• Quality Assurance & Compliance Directors: Seeking immediate pre-audit validation before filing formal Scheme-I applications on the BIS Manakonline portal.
• Automotive EV Engineers & Homologation Officers: Preparing CMVR Form 22 documentation and Conformity of Production (COP) plans for ARAI/ICAT prototype testing.
• Regulatory Auditors & Technical Consultants: Performing rapid statutory gap checks across manufacturing facilities.

User Journey

Screens & Routes in the Frontend

The frontend is a focused Single Page Application (SPA) with zero external routing overhead:

• Root Screen (/): Implemented in [bis-frontend/src/App.jsx:L17-L250](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L17-L250). It renders:
  1. Header with live status badge ("Local Model Active").
  2. Statutory Document Selector (2-column interactive checkbox grid for 8 statutory artifacts).
  3. Conversational Audit Stream (scrollable message history displaying user prompts, bot narrative, and interactive audit cards).
  4. Structured Statutory Audit Cards (target standard, governing QCO order, designated testing agency, computed factory readiness percentage, critical deficit alerts, and one-click CSV export button).
  5. Regulatory Query Bar (interactive prompt input with asynchronous submission handling).

---

0.2 Every Feature Trace

F1: Semantic Standard Identification via Dense Vector RAG

• User Action: The user inputs a generic or technical product term (e.g., "TMT rebars 16mm" or "IS 12269").
• Frontend Flow: [App.jsx:L39-L73](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L39-L73) dispatches an HTTP POST to ${API_URL}/chat.
• Backend Flow: [app.py:L73-L93](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L73-L93):
  1. Validates non-empty query.
  2. Invokes Scikit-Learn classifier [bis_product_department_model_v3.pkl](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L25) to predict administrative department (CED, MTD, FAD, MoRTH_AIS, etc.).
  3. Encodes query into a 384-dimensional dense vector using FastEmbed(model_name="BAAI/bge-small-en-v1.5") ([app.py:L24](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L24)).
  4. Executes nearest neighbor cosine search on FAISS IndexFlatIP index ([app.py:L90](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L90)) with $k=1$.
  5. Resolves top standard entry from [model2_corpus.json](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/model2_corpus.json).
• Data Touched: model2_rag_faiss.index, model2_corpus.json.
• Response Path: Formulates structured payload containing standard name, QCO order, scheme, portal, and statutory citation.

F2: Domain Guardrailing & Out-of-Domain Filter

• User Action: User submits commercial queries or out-of-domain terms (e.g., "crypto token platform", "wooden chairs", "kitchen spoons").
• Frontend Flow: Dispatched to /chat.
• Backend Flow: [app.py:L94-L99](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L94-L99) calls check_domain_guardrail(query, matched_standard, top_score) in [guardrail.py:L3-L42](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/guardrail.py#L3-L42).
  • Tier 1: Explicit Commercial Blacklist: Blocks crypto, stocks, entertainment, unregulated consumer cutlery, uncertified electronics.
  • Tier 2: Instant Regex Pass: Matches exact 3-to-5 digit standard numbers (e.g., 12269, 156, 2062).
  • Tier 3: Statutory Root Matching: Matches keywords (water, cement, steel, battery, toy) if similarity $\ge 0.20$.
  • Tier 4: Dense Semantic Threshold: Requires cosine similarity score $\ge 0.45$.
• Response Path: Rejects out-of-domain queries with a polite regulatory explanation without returning an audit card.

F3: Factory Readiness & Gap Analysis Engine

• User Action: User checks or unchecks compliance artifacts in the frontend grid (selectedDocs).
• Frontend Flow: [App.jsx:L25](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L25) maintains selectedDocs state array; passed in /chat and /audit/export.
• Backend Flow: [gap_engine.py:L21-L59](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/gap_engine.py#L21-L59):
  1. Inspects standard prefix: if "AIS" is present, selects AIS_HOMOLOGATION_CHECKLIST (weights totaling 100); otherwise selects BIS_SCHEME_1_CHECKLIST (weights totaling 100).
  2. Iterates checklist: accumulates weights of fulfilled items; records deficient items, weighting losses, criticality flags (critical=True), and specific remedial action instructions.
  3. Classifies audit readiness:
     • Audit Ready (Minimal Compliance Risk) if $\text{Score} \ge 85%$ and has_critical_gap == False.
     • Conditional Readiness (Minor Gaps Present) if $\text{Score} \ge 60%$.
     • Non-Compliant (Critical Deficits Identified) if $\text{Score} < 60%$ or critical requirements are missing.
• Response Path: Emitted inside audit_data.audit_pack.gap_and_readiness for rendering and CSV generation.

F4: Statutory Audit Pack CSV Export

• User Action: User clicks the "Export CSV" button on an active audit card.
• Frontend Flow: [App.jsx:L75-L93](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L75-L93) calls handleDownloadCsv(m.queryText).
• Backend Flow: [app.py:L140-L152](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L140-L152):
  1. Re-encodes query text, fetches standard via FAISS, re-evaluates gap engine.
  2. Calls generate_csv_report() in [gap_engine.py:L61-L84](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/gap_engine.py#L61-L84) using Python's io.StringIO and csv.writer.
  3. Streams CSV back as PlainTextResponse(media_type="text/csv") with Content-Disposition: attachment.
• Browser Behavior: Browser creates dynamic Object URL blob and initiates automatic download of Audit_<Standard>.csv.

F5: Conversational Officer & Small-Talk Intent Resolver

• User Action: User enters greetings or procedural inquiries ("hello", "who are you", "what can you do").
• Backend Flow: [app.py:L45-L59](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L45-L59) uses regex-free prefix/suffix matching over CONVERSATIONAL_TRIGGERS and returns contextual regulatory onboarding text without invoking heavy vector search.

---

0.3 Every API

Backend Endpoints Table

Method	Path	Auth	Controller Function	Source Location	Purpose	Request Body	Response Type
GET	/	None	health_check()	[app.py:L69-L71](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L69-L71)	Service health check	None	{"status": "live", "service": "..."}
POST	/chat	None	chat_endpoint()	[app.py:L73-L138](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L73-L138)	Primary compliance audit & chat	{"message": str, "documents": List[str]}	JSON with narrative reply & audit pack
POST	/audit/export	None	export_audit()	[app.py:L140-L152](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L140-L152)	Statutory CSV audit pack generation	{"text": str, "documents": List[str]}	text/csv attachment stream

External APIs, Local Models & SDKs

Provider / Tool	Model Name	Where Used in Code	Deployment Type	API Key Required?
FastEmbed (BAAI)	BAAI/bge-small-en-v1.5	[app.py:L24](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L24), [rebuild_faiss_index.py:L7](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/rebuild_faiss_index.py#L7)	Local ONNX Runtime (CPU)	No (Completely Air-Gapped)
Scikit-Learn	bis_product_department_model_v3.pkl	[app.py:L25](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L25)	Local serialized joblib model	No
FAISS	model2_rag_faiss.index	[app.py:L26](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L26)	In-memory IndexFlatIP vector index	No

Environment Variables

Variable Name	Service	Default Value	Used In Code	Description
PORT	Uvicorn Server	10000 (Docker) / 8000 (Local)	[Dockerfile:L4,L15](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/Dockerfile#L4)	Binding port for ASGI HTTP server
PYTHONUNBUFFERED	Python Runtime	1	[Dockerfile:L3](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/Dockerfile#L3)	Prevents stdout buffering for real-time Docker logging

---

0.4 Every Hook, Helper, and Shared Piece

React Hooks (Frontend)

• useState:
  • messages: Array of chat turns and audit cards ([App.jsx:L18](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L18)).
  • input: Current prompt text in the input box ([App.jsx:L24](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L24)).
  • selectedDocs: Array of document IDs checked by user ([App.jsx:L25](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L25)).
  • loading: Boolean flag indicating in-flight inference ([App.jsx:L26](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L26)).
• useRef:
  • chatEndRef: DOM anchor to trigger auto-scroll on new message addition ([App.jsx:L27](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L27)).
• useEffect:
  • Triggers smooth auto-scroll to chatEndRef whenever messages updates ([App.jsx:L29-L31](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L29-L31)).

Backend Utilities & Helpers

• is_conversational(query: str): Detects greeting keywords ([app.py:L45-L47](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L45-L47)).
• resolve_small_talk(query: str): Maps detected conversational intent to pre-cached regulatory responses ([app.py:L49-L59](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L49-L59)).
• check_domain_guardrail(): 4-tier regulatory validation filter ([guardrail.py:L3-L42](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/guardrail.py#L3-L42)).
• evaluate_readiness(): Checklist evaluation against BIS Scheme-I and AIS Homologation ([gap_engine.py:L21-L59](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/gap_engine.py#L21-L59)).
• generate_csv_report(): Generates statutory audit export in RFC 4180 CSV format ([gap_engine.py:L61-L84](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/gap_engine.py#L61-L84)).

---

0.5 Connections & Data Flow

---

0.6 Final Inventory Counts

Metric Category	Verified Count	Exact Code Source References
Frontend Pages / Screens	1	[bis-frontend/src/App.jsx](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx) (SPA root view)
Frontend UI Components	1 (Monolithic clean component)	App in [App.jsx:L17](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/bis-frontend/src/App.jsx#L17) with 4 sub-sections
React Hooks Used	3 built-in hooks	useState (4 instances), useRef (1 instance), useEffect (1 instance)
Backend REST Endpoints	3 endpoints	GET /, POST /chat, POST /audit/export ([app.py:L69,L73,L140](file:///Users/avishjain/avishjain9-repos/21-bis-compliance-engine/app.py#L69))
Backend Core Features	5 numbered features	F1 (RAG), F2 (Guardrail), F3 (Gap Engine), F4 (Audit Export), F5 (Officer)
Machine Learning Models	2 models	FastEmbed BAAI/bge-small-en-v1.5 (ONNX) + Scikit-Learn joblib Classifier
Vector Database Indexes	1 index	FAISS IndexFlatIP (384 dimensions, cosine inner product)
Static Knowledge Records	47 curated standards	Master statutory records across 7 ministries in model2_corpus.json
Backend Middleware	1 middleware	CORSMiddleware (allow origins *, methods *, headers *)
Python Requirements	8 direct packages	fastembed, faiss-cpu, scikit-learn, fastapi, uvicorn, pydantic, python-multipart, numpy
Frontend Dependencies	3 runtime, 4 dev	Runtime: @huggingface/transformers, lucide-react, react, react-dom