At a glance

• Code never changes. Code is hidden before translation and put back letter for letter. If anything does not match, you get an error, never broken code.
• Technical words stay English. A glossary of 1,600 terms in 32 domains keeps array, API or dependency injection as the words you will search for.
• Learn as you read. Click a paragraph to see its English, or read both side by side.
• Works with any AI tool. Select or copy an answer and press Ctrl+Alt+Shift+T (Cmd on Mac), paste it into the panel, or ask @bhashabridge in Copilot Chat.
• Your code never leaves your machine. Only the explanation, with code hidden, goes to the translator you choose: Sarvam AI with your own key, or your Copilot model.

Install: search for BhashaBridge in the VS Code Extensions view, or run ext install SunilKapadia.bhashabridge from the Quick Open box (Ctrl+P / Cmd+P). Then run BhashaBridge: Set Sarvam API Key (or use GitHub Copilot) and press Ctrl+Alt+Shift+T on any AI answer.

Status: 0.1.0 preview. Help is very welcome, especially from Gujarati speakers, and from anyone who wants to add their own language. See Contributing.

More detail: the idea in simple words · technical overview

Why BhashaBridge?

Developers across Gujarat and the rest of India understand loops, functions, APIs, and data structures. But when Copilot or another assistant answers with three paragraphs of dense English explaining why the code is broken, the hard part stops being the bug and becomes the English.

Reading complex sentences in a second language while also reasoning about a technical fix is a heavy cognitive load. It slows learning and quietly tells talented developers that the tools were not built for them.

BhashaBridge removes that barrier. Paste the explanation, pick your language, and read it the way you think.

What it looks like

Input (from any LLM tool):

The useEffect hook runs after every render. Add an empty dependency array so that loadData() is called only once.

Gujarati output:

useEffect hook દરેક render પછી ચાલે છે. ખાલી dependency array ઉમેરો જેથી loadData() ફક્ત એક જ વાર call થાય.

The explanation is translated. useEffect, loadData(), render, dependency array, and call are not, because those are the words you will search for, read in docs, and type in code.

(Illustrative example. Real output depends on the translation provider.)

Principles

How it works

flowchart LR
    A[Paste LLM output] --> B[Markdown parser]
    B -->|code replaced with<br/>SKIP_001, SKIP_002 ...| C[Chunker]
    B -.->|placeholder map| F
    G[(Glossary)] --> D
    C --> D[Translation provider]
    D --> F[Reassembly]
    F --> H[Translated markdown<br/>in the webview]

1. Parse. The input is parsed into a markdown syntax tree. Every code block, inline code span and code line is swapped for a placeholder token such as SKIP_001, and the original is stored in a map.
2. Protect terms. Glossary terms are swapped for placeholders too, so they cannot be translated.
3. Chunk. Paragraphs are grouped into as few requests as fit the provider's limit; a paragraph that is too long is split at sentence ends.
4. Translate. Each request goes to the selected provider with a prompt that keeps technical terms in English. Every reply is checked: each placeholder must come back exactly once.
5. Reassemble. Translations and the exact original code are put back, byte for byte.
6. Render. The result is shown as sanitized HTML in the panel or the chat.

Because code is removed before translation, a model never gets the chance to "helpfully" translate your variable names.

Supported languages and providers

Project status

BhashaBridge 0.1.0 is a preview release on the VS Code Marketplace. See the changelog.

• [x] Milestone 0: Project setup
• [x] Milestone 1: Extension skeleton with the bhashabridge.translate command and webview panel
• [x] Milestone 2: Markdown parser that extracts code into placeholder tokens
• [x] Milestone 3: Translation service (TranslationProvider interface and Sarvam adapter)
• [x] Milestone 4: Reassembly engine
• [x] Milestone 5: Webview UI (input, language and provider pickers, output)
• [x] Milestone 6: Secrets and settings
• [x] Milestone 7: Testing and first release (0.1.0 preview; see docs/PUBLISHING.md)

See docs/TASKS.md for the full breakdown.

Getting started (development)

Requirements: Node.js 22+ and VS Code 1.90+.

git clone https://github.com/SkayKapadia/bhashabridge.git
cd bhashabridge
npm install
npm run bundle

Then open the folder in VS Code and press F5. In the window that opens:

1. Run BhashaBridge: Set Sarvam API Key from the Command Palette (Ctrl+Shift+P / Cmd+Shift+P) and paste a key from dashboard.sarvam.ai. With GitHub Copilot you can skip this and pick VS Code model instead.
2. Run BhashaBridge: Translate LLM Output, paste an AI answer, keep Gujarati selected, and click Translate.

Scripts

Project structure

bhashabridge/
├── src/
│   ├── extension.ts      Entry point: registers commands and the chat participant
│   ├── parser/           Code extraction into placeholders, paragraph segments
│   ├── glossary/         Domain glossaries, term protection, workspace glossary
│   ├── chunking/         Grouping paragraphs into requests
│   ├── translation/      Provider interface, Sarvam and VS Code model adapters, the pipeline
│   ├── reassembly/       Putting translations and code back together
│   ├── webview/          The panel, safe rendering, learning mode
│   ├── chat/             @bhashabridge in the Chat view
│   ├── selection/        Translate Selection shortcut
│   └── settings/, secrets/, cache/, feedback/, quality/
├── media/                The panel's script and styles
├── glossary/domains/     32 domain glossary files
├── test/                 Unit tests, VS Code tests, quality samples
├── scripts/              Quality report and glossary review
└── docs/                 SDD, tasks, overviews, glossary review list

The technical overview explains each part.

Privacy and security

• API keys are stored with VS Code's encrypted SecretStorage. They are never written to settings files, never logged, and never sent anywhere but the provider you chose.
• Your code never leaves your machine. It is replaced by placeholders before anything is sent.
• The explanation text is sent only to the selected translation provider, over HTTPS.
• No telemetry. No analytics, no tracking, no BhashaBridge backend.
• Locked-down webview. A strict Content Security Policy blocks inline and external scripts, and translated output is sanitized before rendering.
• Fully open source, so anyone can audit exactly what happens to their text.

The glossary

Some terms read best in English, and some have a natural local word. The glossary records that decision once, so translations stay consistent. It lives in glossary/domains/, one file per domain:

• Software development (on by default): programming fundamentals, web, backend, databases, cloud and DevOps, security, data and AI, mobile, testing, Git, networking, Agile, embedded and IoT, games, blockchain.
• Business domains (opt-in, for software built in these fields): supply chain, healthcare, finance and banking, insurance, e-commerce, education, manufacturing, government services, telecom, HR and payroll, marketing, legal, real estate, agriculture, energy, travel, media.

Pick domains with BhashaBridge: Choose Glossary Domains. Each term is treated one of three ways:

{
  "keep": ["dependency injection", "callback", "UPI"],
  "annotate": { "ledger": { "gu": "ખાતાવહી", "reviewed": false } },
  "translate": { "inheritance": { "gu": "અનુવંશિકતા (inheritance)", "reviewed": false } }
}

• keep: always stays in English. This is guaranteed: the term is swapped out before translation, like code. Capitals-only terms (IT, PO) match only in capitals.
• annotate: stays in English, with its meaning added after the first use: "ledger (ખાતાવહી)".
• translate: the model is asked to use the given rendering.

Your project's own terms: run BhashaBridge: Create Workspace Glossary to add .bhashabridge/glossary.json to your project, in the same format. Use it for product names, internal acronyms and preferred translations. Its entries win over the built-in glossary, and edits apply to the next translation.

Gujarati renderings are used only once a native speaker marks them "reviewed": true; until then the term stays in English. npm run glossary:review lists the ones still to check in docs/GLOSSARY_REVIEW.md. Think a term is handled badly? Open a pull request.

Roadmap

Contributing

Contributions are very welcome, especially from developers who think in Gujarati or another Indian language. You can help by:

• Adding your language. docs/ADDING_A_LANGUAGE.md walks through it, using Hindi as the example.

• Improving the glossary. Suggest how a technical term should be handled.

• Reporting translation quality. Share an explanation that came out wrong and what it should have said.

• Adding a provider. Implement the TranslationProvider interface for a new backend.

• Testing with real output. Try explanations from Copilot, Continue.dev, or ChatGPT and report what breaks.

See CONTRIBUTING.md for setup, the glossary rules and how reviews work. Every pull request is tested automatically.

License

MIT. Use it, change it, share it.