AI Sheet Cache 🟡 BETA

Intelligent caching for AI-powered spreadsheet formulas

Overview

The AI Sheet Cache provides an asynchronous evaluation pipeline for =BOT_AI_PROMPT() formulas inside spreadsheets. Each formula invocation is cached using a SHA256 hash of the prompt template and referenced cell values, preventing redundant LLM API calls and reducing costs.

Architecture

User edits cell with =BOT_AI_PROMPT()
  │
  ▼
Formula Parser (botsheet-core)
  │
  ▼
  ├── SHA256 cache key (prompt + cell values)
  ├── Cache hit → return cached result immediately
  └── Cache miss →
       ├── Acquire Semaphore (max 10 concurrent)
       ├── Call LLM API
       ├── Store in cache
       └── Return result

BOT_AI_PROMPT Formula

=BOT_AI_PROMPT("Analyze the tone of: ", A2)
=BOT_AI_PROMPT("Translate to Portuguese: ", B2)
=BOT_AI_PROMPT("Summarize in 3 bullet points: ", C2, D2)

The formula parser extracts all cell references as dependencies. When referenced cells change, the cached result is invalidated automatically.

Cache Layer

Batch Evaluation

Batch request format:

{
  "cells": [
    { "prompt": "Analyze: ", "values": ["Great product!"] },
    { "prompt": "Translate: ", "values": ["Hello world"] }
  ]
}

Batch response:

{
  "results": [
    { "index": 0, "value": "Positive feedback", "cached": false },
    { "index": 1, "value": "Olá mundo", "cached": true }
  ]
}

Cost Controls

• Maximum batch size: 100 formulas per request
• Concurrent LLM requests: 10 (configurable via Semaphore)
• Cache TTL: configurable per-bot via config.csv
• Error cells render #VALUE! or #TIMEOUT! — never crash

UI Integration

The Sheet app (see Sheet - Spreadsheets) renders a loading state (...) while AI formulas are being evaluated, then swaps the cell content via HTMX when the result is ready.

Configuration

Feature Flag

Enable with sheet feature flag in Cargo.toml:

botserver = { features = ["sheet"] }