AI Builder - Minecraft AI Assistant

v1.5.0

New Features

• RAG knowledge base retrieval ([KNOWLEDGE] tag) — The AI can now proactively consult the mod's built-in Minecraft knowledge base (blocks, mobs, commands, game mechanics, etc.). When precise information is needed, it names documents via a [KNOWLEDGE]doc name[/KNOWLEDGE] tag to read their body text, instead of relying solely on fixed content hardcoded into the system prompt. Added three new config options: rag_enabled (on by default), rag_max_docs (default 8), and rag_max_chars (default 20000), controlling whether retrieval is enabled and the per-request document count/character limits. Knowledge base docs are released to ai-helper/knowledge/ on first launch (English only, to keep jar size down; the AI can read English docs and still answer in the player's language).
• Structure format conversion (NBT / Litematic / TXT, any direction) — The unified structure browser now has a "Convert" entry point that can batch-convert .nbt, .litematic, and .txt (MCBLUEPRINT v2) files into any target format. Supports selecting multiple source files and a target directory; conversion fully preserves block states, container items (chests, barrels, etc.), and sign front/back text (including legacy pre-1.20 Text1–Text4 compatibility).

Improvements

• Trimmed the system prompt; redstone knowledge now comes from the knowledge base — Removed roughly 500 lines of hardcoded redstone block/circuit reference material from AICommandExecutor, replaced with a pointer telling the AI to consult the knowledge base instead. Also added missing 1.20.4 blocks (TNT, daylight detector, sculk sensor) and a new redstone circuit/logic gate design doc to the knowledge base.
• Display language now automatically follows the game language — Removed the language config option and its manual-switch entry point: the client reads Minecraft's current game language in real time and applies it immediately, while dedicated servers always use English. The "Mod Language Settings" screen is now a static page that only explains this behavior, with no switch buttons.
• Fixed filename sanitization logic — Added ModPaths.sanitizeFileName, which only filters characters truly disallowed by the filesystem (\ / : * ? " < > | and control characters), instead of replacing Chinese, Japanese, and other Unicode characters or spaces with underscores. Applied consistently across AI-generated files, selection exports, and TXT exports.
• Removed hard block-count caps for fill/clear/terrain queries — Replaced with soft prompt-level guidance (fill/clear recommended to stay under ~10,000 blocks, terrain queries under ~30,000 blocks in volume). Exceeding these is no longer rejected with an error; instead the AI is prompted to split the operation into multiple calls, reducing unnecessary failures.
• Removed the teleport tool — The AI-triggerable teleport instruction has been removed, simplifying the set of available actions.
• Refactored the multi-round tool loop — Introduced a unified finishing mechanism (ToolLoopFinisher) ensuring the confirmation flow and network calls (web search, web scraping, resuming the AI) all run on separate threads instead of blocking the server's main tick thread.

Fixes

• Knowledge base content corrections — Fixed the comparator/hopper/dropper container signal strength formula (the previous simplified items/5 approximation was replaced with the standard 1+floor(14×average fill ratio) formula), and fixed an incorrect direction rule for note block instruments (determined by the block below, not above).
• Lost sign text during conversion — Fixed an issue where blank lines in the middle of sign text were mistakenly treated as the end of a section during NBT/Litematic-to-TXT conversion, causing subsequent lines to be lost. The conversion logic also gained support for extracting and writing container items and sign front/back text (including legacy formats).
• Remaining hardcoded Chinese text — Fixed leftover hardcoded Chinese strings in the /aipos command echo and the NBT structure summary; filled in missing keys in zh_cn.json so it fully matches en_us.json.

Tests

• Added KnowledgeBaseTest, NbtToTxtConverterTest, StructureFormatConverterTest, and LitematicParserTest unit tests, covering knowledge base retrieval, structure format conversion, and Litematic parsing.

Security & Permission Model Changes

• execute_command no longer executes with elevated server-side privileges — Vanilla commands suggested by the AI are no longer executed directly on the server with OP permissions. Instead, they are sent to the requesting player's client, which automatically opens the chat screen with the command pre-filled. Whether the command is sent, and with what permission level, is entirely up to the player, removing the risk of "remote backend control of privileged command execution." The server-side command blacklist has been removed accordingly, since the server no longer executes commands directly.
• New standalone AI Permission Settings screen — Split out from AI Chat Settings into K → AI Permission Settings, centralizing the following permission-related options:
  • Allow AI to Use Vanilla Commands (vanilla_commands_enabled, on by default): when off, the system prompt no longer includes execute_command instructions, so the AI can only use the mod's built-in features; even if the AI still tries to generate that instruction, it is rejected outright at runtime.
  • Require Confirmation Before Execution (confirm_before_execute_enabled, off by default): when on, before the AI uses a mod-specific feature (placing blocks, building blueprints, terrain queries, web search, web scraping, etc.) it first sends a [Yes]/[No] confirmation message in chat. Multiple actions within the same round are merged into a single batch confirmation, and unconfirmed requests are automatically canceled after 60 seconds.
  • Max tool-call rounds (the max_tool_rounds setting, moved here from the chat settings screen).
• New /aiconfirm command — Handles [Yes]/[No] button clicks in chat for pending AI action requests. Once confirmed, the conversation automatically resumes, covering all tool types (web search, web scraping, terrain queries, in-game actions), preventing the conversation from stalling or actions from being silently dropped after confirmation.

Documentation

• Synced the README, user manuals (Chinese & English), and in-game manuals with the latest behavior: updated the execute_command permission model description, added documentation for the new permission settings screen, the /aiconfirm command, and new config options; removed outdated teleport references; bumped the version number to 1.4.2.