Bharat Bricks Hacks 2026 · IIT Delhi Built on Databricks · Powered by Meta Llama 4 Maverick · Indian Law RAG

The motivation behind NyaaAIk is to democratize the accessibility and affordability of legal resources in India. Navigating the Indian judicial system can be daunting and prohibitively expensive for the average citizen.

NyaaAIk is a Databricks-native legal assistant using Llama 4 Maverick and Sarvam Saaras v3. It employs a state-of-the-art Hybrid RAG architecture: Databricks Genie autonomously chunks statutory laws into Delta Tables, synced natively with Vector Search. Simultaneously, an LLM refines user queries for parallel semantic vector retrieval and live web scraping of recent rulings. This bridges the gap—providing courtroom-grade insight to advocates while translating complex statutory laws into actionable, plain-language guidance for everyday citizens.

NyaaAIk is an AI legal assistant for Indian users that:

• Searches the Bharatiya Nyaya Sanhita (BNS), BNSS, and BSA using RAG
• Finds live court judgments from Indian Kanoon
• Understands voice in Hindi, Tamil, Bengali, Telugu and 7 more Indian languages (Sarvam Saaras v3)
• Analyzes uploaded PDFs, Word docs, and screenshots using Llama 4 Maverick vision
• Responds in two modes: Advocate (courtroom English) and Citizen (plain language)
• Works in Dark and Light themes

flowchart TD
    %% Base Inputs
    U[User Query] --> Refiner[LLM Query Refiner<br/>Llama 4 Maverick]

    %% Ingestion & Vector Search Pipeline
    subgraph DataPlatform[Databricks Data Intelligence Platform]
        direction TB
        BNS[(Delta Tables<br/>BNS / BNSS / BSA)]
        FallBk["/Planned: Chunk Set 2<br/>2000 Case Fallback<br/>Not done — data unavailable in cluster"/]

        Genie([Databricks Genie<br/>Auto-Chunking]) --> BNS
        Genie -.-> FallBk

        BNS --> VS1[(Databricks Vector Search<br/>Index 1)]
        FallBk -.-> VS2["/Planned: Vector Search Index 2<br/>Not done — data unavailable in cluster"/]
    end

    %% Live Searching
    subgraph WebScraping[Live Web Scraping]
        Kanoon([Indian Kanoon API])
    end

    %% Retrieval Layer
    Refiner -- Semantic Query --> VS1 & VS2
    Refiner -- Refined Keywords --> Kanoon

    VS1 -- Highly Relevant Statutes --> Synth
    VS2 -.->|Data Unavailable| Synth
    Kanoon -- Top 5-10 Live Cases --> Synth

    %% Synthesis Layer
    subgraph LLM[LLM Synthesis Engine]
        direction TB
        Persona["UI Prompts/Modifiers<br/>Advocate vs Citizen"] --> Synth["Meta Llama 4 Maverick<br/>Open Source LLM"]
        U -- Original Context --> Synth
    end

    %% Output
    Synth --> Out["Structured Legal Response<br/>+ Properly Cited Cases"]

    classDef db fill:#ff3621,stroke:#333,stroke-width:1px,color:#fff;
    classDef os fill:#1a73e8,stroke:#333,stroke-width:1px,color:#fff;
    classDef planned fill:#f0f0f0,stroke:#666,stroke-width:2px,stroke-dasharray: 5 5,color:#666;

    class VS1,Refiner db;
    class Synth os;
    class FallBk,VS2 planned;

Note: Highlighted components leverage the open-source Meta Llama 4 Maverick model and extensive Databricks enterprise infrastructure.

The entire data pipeline that powers NyaaAIk's legal knowledge base is run through two Databricks notebooks, executed in sequence inside the workspace. The diagram below shows exactly what each notebook does and how data flows end-to-end from raw law files to the deployed app.

flowchart TD
    classDef notebook fill:#1a3c6e,stroke:#4a90d9,stroke-width:2px,color:#fff
    classDef delta fill:#cc3300,stroke:#ff6644,stroke-width:2px,color:#fff
    classDef volume fill:#2d6a4f,stroke:#52b788,stroke-width:2px,color:#fff
    classDef app fill:#1a5c1a,stroke:#52b788,stroke-width:2px,color:#fff
    classDef source fill:#444,stroke:#999,stroke-width:1px,color:#fff

    %% ── RAW DATA SOURCES ─────────────────────────────────────────────────
    subgraph Sources["Raw Data Sources"]
        CSV["BNS Sections CSV<br/>Volume upload or GitHub mirror"]
        PDF["Official Law PDFs<br/>BNS / BNSS / BSA / Constitution<br/>MHA Gazette / Indiacode.nic.in"]
        HF["Hugging Face Dataset<br/>IPC to BNS Mapping"]
    end

    %% ── NOTEBOOK 1 ───────────────────────────────────────────────────────
    subgraph NB1["NOTEBOOK 1 — india_legal_policy_ingest.ipynb"]
        direction TB
        N1_0["Step 0: Install Dependencies<br/>requests pandas bs4 pymupdf openpyxl"]
        N1_1["Step 1: Config and Helpers<br/>Read secrets from nyaya-dhwani scope<br/>CATALOG=main / SCHEMA=india_legal"]
        N1_1b["Step 1b: Sarvam Connectivity Test<br/>Optional ping to sarvam.ai"]
        N1_2["Step 2: Unity Catalog Setup<br/>CREATE DATABASE main.india_legal<br/>CREATE VOLUME legal_files"]
        N1_3["Step 3: Load BNS Sections<br/>1 Try Volume CSV / 2 GitHub mirrors<br/>3 PDF + ai_parse_document fallback"]
        N1_4["Step 3c: Download Official Law PDFs<br/>BNS / BNSS / BSA / Constitution<br/>Saved to Unity Catalog Volume"]
        N1_5["Step 4: BNS to IPC Mapping Table<br/>Fetch from Hugging Face<br/>Fallback: 7-row built-in stub"]
        N1_6["Step 5: Build legal_rag_corpus<br/>Merge BNS + IPC mapping chunks<br/>Save as Delta Table"]
        N1_7["Step 6: Verify Tables<br/>SHOW TABLES IN main.india_legal<br/>SELECT preview from legal_rag_corpus"]

        N1_0 --> N1_1 --> N1_1b --> N1_2 --> N1_3 --> N1_4 --> N1_5 --> N1_6 --> N1_7
    end

    %% ── DELTA TABLES ─────────────────────────────────────────────────────
    subgraph DeltaLayer["Delta Lake — main.india_legal"]
        T1["bns_sections<br/>Delta Table"]
        T2["bns_ipc_mapping<br/>Delta Table"]
        T3["legal_rag_corpus<br/>chunk_id / source / doc_type / title / text"]
        T4["bns_parsed_elements<br/>Optional — PDF ai_parse_document path"]
    end

    %% ── NOTEBOOK 2 ───────────────────────────────────────────────────────
    subgraph NB2["NOTEBOOK 2 — build_rag_index.ipynb"]
        direction TB
        N2_0["Cell 1: Set REPO_ROOT<br/>Edit path to your workspace clone"]
        N2_1["Cell 1 run: Install RAG Stack<br/>numpy / faiss-cpu 1.7.x / sentence-transformers<br/>pip install -e repo rag rag_embed"]
        N2_2["Cell 2: Load Corpus from Delta<br/>SELECT from legal_rag_corpus to Pandas"]
        N2_3["Cell 3: Embed All Chunks<br/>all-MiniLM-L6-v2 / normalize / float32"]
        N2_4["Cell 4: Save RAG Artifacts<br/>corpus.faiss / chunks.parquet / manifest.json<br/>to nyaya_index Volume"]
        N2_5["Cell 5: Smoke Test<br/>Load index / search theft under BNS / top 5"]

        N2_0 --> N2_1 --> N2_2 --> N2_3 --> N2_4 --> N2_5
    end

    subgraph NB3["NOTEBOOK 3 — setup_vector_search.py (Optional)"]
        direction TB
        N3_1["Create VS Endpoint: nyaya_vs_endpoint<br/>STANDARD type / 20-50ms latency"]
        N3_2["Enable CDF on legal_rag_corpus<br/>Required for Delta Sync index"]
        N3_3["Create Delta Sync Index<br/>legal_rag_corpus_index<br/>Embedding: databricks-bge-large-en"]
        N3_4["Trigger Sync and Wait for ONLINE<br/>Grant CAN_QUERY to App service principal"]

        N3_1 --> N3_2 --> N3_3 --> N3_4
    end

    subgraph NB4["NOTEBOOK 4 — run_benchmark.py (Evaluation)"]
        direction TB
        B1["Phase 1: Retrieval Eval<br/>FAISS vs Vector Search<br/>Recall@7 and MRR"]
        B2["Phase 2: MCQ Accuracy<br/>Internal + BhashaBench-Legal<br/>English and Hindi"]
        B3["Phase 3: Open-Ended Quality<br/>Keyword coverage / chunk recall"]
        B4["Phase 4: Multilingual<br/>Sarvam translation round-trip<br/>Save results to Delta"]

        B1 --> B2 --> B3 --> B4
    end

    %% ── VOLUME OUTPUT ────────────────────────────────────────────────────
    subgraph VolumeOut["Unity Catalog Volume — nyaya_index"]
        V1["nyaya_index/<br/>corpus.faiss — FAISS index<br/>chunks.parquet — metadata<br/>manifest.json — model info"]
    end

    %% ── DEPLOYED APP ─────────────────────────────────────────────────────
    subgraph AppLayer["Deployed NyaaAIk — Databricks Apps"]
        Retriever["retriever.py<br/>Vector Search + FAISS fallback"]
        Flask["Flask API app/main.py<br/>/api/chat / /api/transcribe / /api/upload"]
        React["React Frontend<br/>Advocate / Citizen / Voice / Upload"]
    end

    %% ── DATA FLOW ────────────────────────────────────────────────────────
    CSV --> N1_3
    HF  --> N1_5
    PDF --> N1_4

    N1_3 --> T1
    N1_5 --> T2
    N1_6 --> T3
    N1_3 -.->|PDF path only| T4

    T3 --> N2_2
    N2_4 --> V1

    T3 -.->|Optional VS setup| N3_2
    N3_3 --> VS_IDX[("legal_rag_corpus_index<br/>Databricks Vector Search")]

    V1  --> Retriever
    VS_IDX --> Retriever
    T3  --> Retriever
    Retriever --> Flask --> React

    V1 -.->|Evaluation| B1
    T3 -.->|Evaluation| B1

    %% ── CLASS ASSIGNMENTS ────────────────────────────────────────────────
    class NB1,NB2,NB3,NB4 notebook
    class T1,T2,T3,T4,VS_IDX delta
    class V1 volume
    class Retriever,Flask,React app
    class CSV,PDF,HF source

Follow these steps in order inside your Databricks workspace. No local Python needed for data setup.

Before running any notebook, ensure:

• Databricks workspace with Unity Catalog enabled
• Secret scope nyaya-dhwani created (see Deployment Step 4) containing:
  • sarvam_api_key
  • indian_kanoon_api_token
  • datagov_api_key (optional)
• Repo cloned into Workspace → Repos (see Step 8)
• A compute cluster or Serverless notebook attached

This error means either:

• Unity Catalog is not fully configured in your workspace, or
• The main catalog doesn't exist yet, or
• You are on a trial workspace that uses a different default catalog name

Fix Option A — Create the catalog and schema manually (recommended):

Open any Databricks notebook, attach to a cluster, and run this before Notebook 1:

-- Run this in a SQL cell or notebook
CREATE CATALOG IF NOT EXISTS main;
CREATE SCHEMA  IF NOT EXISTS main.india_legal;
CREATE VOLUME  IF NOT EXISTS main.india_legal.legal_files;

Or via the UI: Catalog → + Add → Add a catalog → name it main → Create.

Fix Option B — Use your own catalog name:

If your workspace has a different catalog (e.g. hive_metastore or a custom one), open Notebook 1 and edit Step 1 to change:

CATALOG  = 'main'        # ← change to your catalog name, e.g. 'hive_metastore'
SCHEMA   = 'india_legal' # ← keep as-is or rename
VOLUME   = 'legal_files' # ← keep as-is or rename

⚠️ If you change CATALOG, also update app.yaml and setup_vector_search.py to use the same name consistently.

Verify Unity Catalog is active:

SHOW CATALOGS;
-- Should list 'main' and your workspace catalog

Do this before running Notebook 1. The notebooks need the raw legal text data to exist somewhere accessible. You have two paths:

1. Download the BNS sections CSV from Kaggle:
   👉 kaggle.com/datasets/nandr39/bharatiya-nyaya-sanhita-dataset-bns

2. In your Databricks workspace, go to:
   Catalog → main → india_legal → Volumes → legal_files → Upload to this volume

3. Upload bns_sections.csv (rename it exactly to bns_sections.csv if needed)

4. Notebook 1 → Step 3 will automatically find it at /Volumes/main/india_legal/legal_files/bns_sections.csv

Databricks Genie can read raw PDFs, CSVs, and text directly from your Volume and convert them into structured Delta tables automatically.

Step-by-step:

1. Upload raw law files to the Volume:

   • Go to Catalog → main → india_legal → Volumes → legal_files → Upload
   • Upload any of the following you have:
     • bns_sections.csv — BNS section text (structured)
     • bns_2023.pdf — Official BNS Gazette PDF (MHA)
     • bnss_2023.pdf — BNSS Gazette PDF
     • bsa_2023.pdf — BSA Gazette PDF
     • constitution_of_india.pdf — Constitution PDF

2. Open Databricks Genie:

   • Go to New → Genie Space in the left sidebar
   • Or navigate to Genie from the sidebar

3. Tell Genie what to do — example prompts you can use:

   I have uploaded bns_sections.csv to /Volumes/main/india_legal/legal_files/.
   Please read it and save it as a Delta table called main.india_legal.bns_sections
   with columns: section_number, section_title, section_text, source
   

   Read the PDF at /Volumes/main/india_legal/legal_files/bns_2023.pdf,
   extract all sections with their numbers and text,
   and save as Delta table main.india_legal.bns_sections
   

4. Verify Genie created the table:

   SELECT COUNT(*) FROM main.india_legal.bns_sections;
   SELECT * FROM main.india_legal.bns_sections LIMIT 5;

5. Once the table exists, Notebook 1 Step 5 (Build Corpus) will pick it up automatically when building legal_rag_corpus

💡 Tip: Genie works best with structured CSVs. For PDFs, the notebook's ai_parse_document fallback in Step 3b does the same thing automatically — so if Genie struggles with a PDF, just upload the file to the volume and let Notebook 1 handle it.

💡 Tip: If you don't have any files yet — Notebook 1 Step 3 will automatically download the official PDFs from the MHA Gazette and extract sections for you. You don't need to upload anything manually in that case.

File: notebooks/india_legal_policy_ingest.ipynb

Open this notebook in your Databricks workspace. Run each cell in order:

Expected output after this notebook:

✅ Schema : main.india_legal
✅ Volume : /Volumes/main/india_legal/legal_files
🏛️  legal_rag_corpus: <N> total chunks
  BNS_2023         → <N> rows
  BNS_IPC_MAPPING  → <N> rows

💡 Tip — Manual CSV upload: If Step 3 fails to find a BNS CSV, download one from
Kaggle: Bharatiya Nyaya Sanhita Dataset,
upload via Catalog → Volumes → legal_files → Upload to this volume, then re-run Step 3.

File: notebooks/build_rag_index.ipynb

⚠️ Run this notebook AFTER Notebook 1 completes — it reads from legal_rag_corpus.

Open this notebook in Databricks. Run each cell in order:

Expected output:

✅ pip: numpy 1.x, pandas<3, faiss-cpu 1.7.x, pyarrow, sentence-transformers
✅ import nyaya_dhwani → /Workspace/Users/.../src/nyaya_dhwani
✅ import faiss OK
(<N_chunks>, 384)   ← embedding dimensions
✅ Artifacts saved to /Volumes/main/india_legal/legal_files/nyaya_index/

Output artifacts saved to:

/Volumes/main/india_legal/legal_files/nyaya_index/
├── corpus.faiss      ← FAISS similarity index
├── chunks.parquet    ← chunk metadata (id, source, title, text)
└── manifest.json     ← embedding model name, catalog, schema, timestamp

💡 If import nyaya_dhwani fails: Ensure REPO_ROOT exactly matches the path shown in the Workspace sidebar. Do not add %restart_python in the install cell — it breaks the kernel before later cells run.

The retriever.py module in the deployed app automatically loads the FAISS index from the Volume path. No additional steps are needed — just deploy the app as described in Steps 9–10 below.

nyaya-dhwani-hackathon/
│
├── app/
│   ├── main.py              ← Flask backend (all API routes)
│   └── static/              ← Built React app (DO NOT edit manually)
│       ├── index.html
│       └── assets/
│
├── notebooks/
│   ├── india_legal_policy_ingest.ipynb  ← NOTEBOOK 1: Data ingestion → Delta tables
│   ├── build_rag_index.ipynb            ← NOTEBOOK 2: Build FAISS index → Volume
│   ├── setup_vector_search.py           ← NOTEBOOK 3: Create VS endpoint + Delta Sync index (optional)
│   └── run_benchmark.py                 ← NOTEBOOK 4: RAG evaluation — retrieval / MCQ / multilingual
│
├── frontend/
│   ├── src/
│   │   ├── App.jsx          ← Root component + theme + state
│   │   ├── index.css        ← All styles (dark + light themes)
│   │   ├── components/
│   │   │   ├── Topbar.jsx        ← Header + theme toggle
│   │   │   ├── Sidebar.jsx       ← Chat history + docs list
│   │   │   ├── ControlsBar.jsx   ← Persona + Court + Style controls
│   │   │   ├── TopicsChips.jsx   ← Quick topic selector
│   │   │   ├── WelcomeScreen.jsx ← First-load suggestions
│   │   │   ├── ChatWindow.jsx    ← Message thread
│   │   │   ├── InputBar.jsx      ← Text input + mic + upload
│   │   │   └── UploadModal.jsx   ← Document upload UI
│   │   ├── hooks/
│   │   │   ├── useChat.js        ← Chat state + API calls
│   │   │   └── useSarvamVoice.js ← MediaRecorder → Sarvam STT
│   │   └── utils/
│   │       ├── api.js            ← All fetch calls to Flask
│   │       └── icons.jsx         ← SVG icon components
│   ├── package.json
│   └── vite.config.js
│
├── src/
│   └── nyaya_dhwani/
│       ├── sarvam_client.py  ← Sarvam API helpers (STT, TTS, translate)
│       ├── llm_client.py     ← Maverick LLM calls
│       ├── retriever.py      ← Databricks Vector Search + FAISS fallback
│       └── case_search.py    ← Indian Kanoon API integration
│
├── app.yaml                  ← Databricks App config
├── setup_secrets.py          ← Script to create secret scope
└── README.md

Follow every step in order. Do not skip.

Before starting, make sure you have:

• A Databricks workspace (Community Edition will NOT work — you need Databricks Apps access)
• Access to Databricks AI Gateway with Llama 4 Maverick configured
• A Vector Search endpoint with the BNS legal corpus indexed
• An Indian Kanoon API token → get it from indiankanoon.org
• A Sarvam API key → get it from dashboard.sarvam.ai
• Git installed locally
• Node.js 18+ installed locally (for building the frontend)

git clone https://github.com/gaganTakIITD/NyaaAIk.git
cd NyaaAIk

Open app.yaml. Update the values to match your Databricks workspace:

command: ["python", "app/main.py"]

env:
  - name: LLM_OPENAI_BASE_URL
    value: "https://<YOUR-GATEWAY-ID>.ai-gateway.cloud.databricks.com/mlflow/v1"

  - name: LLM_MODEL
    value: "databricks-llama-4-maverick"

  - name: NYAYA_RETRIEVAL_BACKEND
    value: "vector_search"

  - name: NYAYA_VS_ENDPOINT_NAME
    value: "nyaya_vs_endpoint"         # ← your Vector Search endpoint name

  - name: NYAYA_VS_INDEX_NAME
    value: "main.india_legal.legal_rag_corpus_index"  # ← your index

  - name: SARVAM_STT_MODEL
    value: "saaras:v3"

⚠️ Do NOT put API keys here. They are loaded from Databricks Secrets (Step 4).

This stores API keys securely. Run the following in a Databricks notebook (not locally):

Cell 1 — Get workspace context

import requests

ctx   = dbutils.notebook.entry_point.getDbutils().notebook().getContext()
HOST  = ctx.apiUrl().get()
TOKEN = ctx.apiToken().get()
HEADERS = {"Authorization": f"Bearer {TOKEN}"}
print("Host:", HOST)

Cell 2 — Create the secret scope

r = requests.post(f"{HOST}/api/2.0/secrets/scopes/create",
    headers=HEADERS,
    json={"scope": "nyaya-dhwani", "initial_manage_principal": "users"})
print(r.status_code, r.text)
# 200 = created   |   RESOURCE_ALREADY_EXISTS = already exists (both OK)

Cell 3 — Store Indian Kanoon API token

r = requests.post(f"{HOST}/api/2.0/secrets/put",
    headers=HEADERS,
    json={
        "scope": "nyaya-dhwani",
        "key":   "indian_kanoon_api_token",
        "string_value": "YOUR_INDIAN_KANOON_TOKEN_HERE"
    })
print(r.status_code, r.text)

Cell 4 — Store Sarvam API key

r = requests.post(f"{HOST}/api/2.0/secrets/put",
    headers=HEADERS,
    json={
        "scope": "nyaya-dhwani",
        "key":   "sarvam_api_key",
        "string_value": "YOUR_SARVAM_API_KEY_HERE"
    })
print(r.status_code, r.text)

Cell 5 — Verify

r = requests.get(f"{HOST}/api/2.0/secrets/list?scope=nyaya-dhwani", headers=HEADERS)
print(r.json())
# Expected: {'secrets': [{'key': 'indian_kanoon_api_token', ...}, {'key': 'sarvam_api_key', ...}]}

✅ Values always show as [REDACTED] in Databricks — that is correct behaviour.

In the Databricks Apps UI, before deploying, add both secrets as resources so the app's service principal can read them:

1. Go to Compute → Apps → (your app) → Resources
2. Add first resource:
   • Secret scope: nyaya-dhwani
   • Secret key: indian_kanoon_api_token
   • Permission: Can read
   • Resource key: secret1
3. Add second resource:
   • Secret scope: nyaya-dhwani
   • Secret key: sarvam_api_key
   • Permission: Can read
   • Resource key: secret2

⚠️ Without these resource declarations, the app cannot read the secrets even if they exist in the scope.

1. Navigate to Workspace → Repos → NyaaAIk → notebooks
2. Open india_legal_policy_ingest.ipynb
3. Attach to a cluster (Standard or Serverless)
4. Run all cells in order — see the Notebook 1 guide above

Run this after Notebook 1 completes successfully.

1. Navigate to notebooks → build_rag_index.ipynb
2. Edit REPO_ROOT in Cell 1 to match your workspace path
3. Run all cells in order — see the Notebook 2 guide above

The built files must be committed to git. Run this on your local machine (not Databricks):

cd frontend
npm install
npm run build
cd ..

This outputs compiled files to app/static/. Verify:

app/static/index.html          ← must exist
app/static/assets/index-*.js   ← must exist
app/static/assets/index-*.css  ← must exist

⚠️ Never run npm run build on Databricks — Node.js is not available there.

git add -A
git commit -m "deploy: update config and frontend build"
git push origin main

1. In your Databricks workspace, go to Workspace → Repos → Add Repo
2. URL: https://github.com/gaganTakIITD/NyaaAIk.git
3. Click Create Repo
4. After creation, click Pull to get the latest code

1. Go to Compute → Apps → Create App
2. Choose Custom App
3. Set Source to the repo you added in Step 8
4. Set App file: app.yaml
5. Add the two secret resources (Step 5)
6. Click Deploy

The app starts. Watch the Logs tab for:

INFO - Indian Kanoon API: configured
INFO - Sarvam STT (Saaras): configured
INFO - Starting NyaaAIk on 0.0.0.0:8000

If you see NOT configured for any key → re-check Steps 4 and 5.

Click the URL shown in the Apps dashboard. You should see the NyaaAIk interface.

Every time you change code locally:

# If you changed frontend code (React/CSS):
cd frontend
npm run build
cd ..

# Always:
git add -A
git commit -m "your message"
git push origin main

# Then in Databricks:
# Workspace → Repos → NyaaAIk → Pull
# Compute → Apps → NyaaAIk → Deploy (or Restart)

• Click the mic button → select language → speak → stop
• Supported: Hindi, English, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Malayalam, Punjabi
• Audio is recorded in browser, sent to /api/transcribe, processed by Sarvam Saaras v3

• PDF: text extracted with pypdf
• DOCX: text extracted with python-docx
• Images / Screenshots: sent as base64 to Llama 4 Maverick vision API
• Ctrl+V: paste a screenshot directly into the input box
• Documents are session-scoped (in server memory) — re-upload after app restarts

• Dark: #131314 background, pure blue #4dabf7 accent
• Light: #ffffff background, Google blue #1a73e8 accent
• Toggle: ☀️/🌙 button top-right — saved to browser localStorage

To run locally for testing UI changes (no Databricks connection):

# Terminal 1 — backend (will error on Databricks-specific imports, use mock mode)
cd app
python main.py

# Terminal 2 — frontend dev server with proxy
cd frontend
npm run dev

Open http://localhost:5173

Note: Vector search and live cases won't work locally without Databricks credentials.

Built for Bharat Bricks Hacks 2026 · IIT Delhi · Team NyaaAIk