SwiftDBAI

A Swift package that adds a natural language query interface to any SQLite database in your iOS, macOS, or visionOS app. Works with GRDB, Core Data, SwiftData, or raw SQLite -- any .sqlite file on the device.

Real-World Demo: NetNewsWire

SwiftDBAI added to NetNewsWire (the open source RSS reader) in one commit. The LLM queries the article database directly -- unread counts, feed rankings, article search. Running Ollama locally, no backend.

iPhone	iPad

Quick Start

Point it at your database, give it an LLM:

import SwiftDBAI
import AnyLanguageModel

DataChatView(
    databasePath: documentsURL.appendingPathComponent("app.db").path,
    model: OllamaLanguageModel(model: "qwen3-coder")
)

Or pass an existing GRDB connection:

DataChatView(
    database: myDatabasePool,
    model: OpenAILanguageModel(apiKey: key, model: "gpt-4o"),
    allowlist: .readOnly,
    additionalContext: "E-commerce app with orders, products, and customers."
)

Schema introspection is automatic. No annotations, no mapping files. Read-only by default.

Tool Calling

If your app already talks to an LLM, use DatabaseTool to register SwiftDBAI as a tool it can call. Your existing LLM generates SQL, SwiftDBAI validates and executes it.

let tool = try await DatabaseTool(databasePath: myDB)

// Add schema to your system prompt
let systemPrompt = "You are a helpful assistant.\n\n" + tool.systemPromptSnippet

// Register with your LLM
let functionDef = tool.openAIFunctionDefinition

// When the LLM calls the tool
let result = try tool.execute(sql: "SELECT * FROM users WHERE active = 1")
result.jsonString     // return to LLM as tool response
result.markdownTable  // display to user

SQL is validated against a read-only allowlist before execution. INSERT, UPDATE, DELETE, and DROP are rejected.

Headless / Programmatic Use

Use ChatEngine directly when you don't need a UI:

let pool = try DatabasePool(path: myDB)
let engine = ChatEngine(
    database: pool,
    model: OpenAILanguageModel(apiKey: key, model: "gpt-4o")
)

let response = try await engine.send("How many users signed up this week?")
print(response.summary)   // "There were 42 new signups this week."
print(response.sql)        // Optional("SELECT COUNT(*) FROM users WHERE ...")

Installation

dependencies: [
    .package(url: "https://github.com/krishkumar/SwiftDBAI.git", from: "1.0.0"),
]

Presentation

Works in every context: sheets, full-screen covers, navigation pushes, UIKit modals.

Custom theme	Sheet presentation

// Sheet with nav bar and Done button
.dataChatSheet(isPresented: $showChat, databasePath: path, model: myLLM)

// Full-screen cover
.dataChatFullScreen(isPresented: $showChat, databasePath: path, model: myLLM)

// UIKit
let vc = DataChatViewController(databasePath: path, model: myLLM)
present(vc, animated: true)

See THEMING.md for customization -- colors, fonts, dark mode, compact mode, assistant avatars.

Features

• Drop-in SwiftUI chat view (DataChatView) or headless ChatEngine
• LLM-agnostic via AnyLanguageModel -- OpenAI, Anthropic, Gemini, Ollama, llama.cpp, or any OpenAI-compatible endpoint
• DatabaseTool for integrating into your existing LLM tool calling setup
• Automatic schema introspection -- no manual annotations
• Safety-first: read-only by default, operation allowlists, table-level mutation policies
• Auto-generated bar charts, pie charts, and data tables
• Structured output via @Generable for models that support it
• 448 tests

Choosing a Provider

// Cloud
ProviderConfiguration.openAI(apiKey: "sk-...", model: "gpt-4o")
ProviderConfiguration.anthropic(apiKey: "sk-ant-...", model: "claude-sonnet-4-20250514")

// Local -- data stays on device
ProviderConfiguration.ollama(model: "llama3.2")
ProviderConfiguration.llamaCpp(model: "default")

// Any OpenAI-compatible endpoint
ProviderConfiguration.openAICompatible(apiKey: "key", model: "llama-3.1-70b", baseURL: url)

Safety

Preset	Allowed
.readOnly (default)	SELECT
.standard	SELECT, INSERT, UPDATE
.unrestricted	SELECT, INSERT, UPDATE, DELETE

Per-table control via MutationPolicy. Destructive operations require confirmation through ToolExecutionDelegate.

Architecture

User Question
    |
    v
ChatEngine
    |-- SchemaIntrospector   (auto-discovers tables, columns, keys, indexes)
    |-- PromptBuilder        (builds LLM system prompt with schema context)
    |-- LanguageModel        (generates SQL via AnyLanguageModel)
    |-- SQLQueryParser       (parses and validates against allowlist/policy)
    |-- GRDB                 (executes SQL against SQLite)
    |-- TextSummaryRenderer  (summarizes results via LLM)
    v
ChatResponse { summary, sql, queryResult }

Demo App

The demo app is at Example/SwiftDBAIDemo/. It includes a showcase of all themes and presentation modes. Generate the Xcode project with xcodegen:

cd Example/SwiftDBAIDemo && xcodegen generate

Requirements

• iOS 17.0+ / macOS 14.0+ / visionOS 1.0+
• Swift 6.1+
• Xcode 16+

License

MIT. See LICENSE for details.