Kalara

Real-time intelligence overlay. Listens to video calls via Recall.ai, RAGs against your docs, surfaces answers as a live floating overlay.

Architecture

Recall.ai bot (joins call)
    → real-time transcript webhook
        → local Express server (port 3847)
            → question detection
            → RAG against your docs (keyword search, upgradeable to embeddings)
            → Claude Sonnet for answer generation
                → Electron IPC
                    → Top bar overlay (auto-hide, appears on question detection)
                    → Sidebar (awareness mode, always-on)

Setup

1. Install dependencies

npm install

2. Configure environment

cp .env.example .env
# Fill in ANTHROPIC_API_KEY, RECALL_API_KEY, WEBHOOK_URL

3. Expose local webhook (for Recall.ai to reach you)

# Install ngrok if you don't have it
brew install ngrok
ngrok http 3847
# Copy the https URL into WEBHOOK_URL in .env

4. Run

npm run dev

5. Upload your pitch docs

• Click the Kalara menubar icon
• Select "Upload Documents"
• Drag in your pitch deck PDF, FAQ doc, objection handling doc

6. Start a session

• Click "Start Recall Session"
• Paste the meeting link (Google Meet, Zoom, Teams all work)
• The Recall.ai bot joins silently

Modes

Default (passive): App is invisible. When a question is detected, the top bar slides down with an answer, auto-hides after 12 seconds.

Awareness Mode: Toggle from menubar. Sidebar opens showing live transcript + answer history. You can also type manual queries into the sidebar search bar.

Upgrading the search

The current implementation uses keyword matching (fast, no cost, good enough for MVP). To upgrade to semantic search:

1. Add OpenAI embeddings (or use Claude's embeddings via Voyage AI)
2. Replace the ragQuery keyword scoring with cosine similarity against embedded chunks
3. ChromaDB is already wired in for this — just needs the embedding step added

Recall.ai webhook events

The server listens for:

• transcript.partial_update — real-time word-level updates
• transcript.word — individual word events

Adjust in src/server/webhook.js based on which Recall plan you're on.