MakerAI Suite v3.8 — The AI Ecosystem for Delphi
🌐 Official Website: https://makerai.cimamaker.com
📖 Manual: https://www.gustavoenriquez.com/book-makerai — available in English and Spanish
> Free Pascal / Lazarus port available — Full port of MakerAI Suite for FPC 3.2+ (12 LLM drivers, RAG, Agents, MCP, Embeddings). See the fpc branch.
MakerAI is more than an API wrapper
Most AI libraries for Delphi stop at wrapping REST calls. MakerAI is different.
Yes, MakerAI includes native, provider-specific components that give you direct, full-fidelity access to each provider's API — every model parameter, every response field, every streaming event, exactly as the provider defines it.
But on top of that, MakerAI is a complete AI application ecosystem that lets you build production-grade intelligent systems entirely in Delphi:
- RAG pipelines (vector and graph-based) with SQL-like query languages (VQL / GQL)
- Autonomous Agents with graph orchestration, checkpoints, and human-in-the-loop approval
- MCP Servers and Clients — expose or consume tools using the Model Context Protocol (dual-era: stateless spec 2026-07-28 + legacy handshake)
- Native ChatTools — bridge AI reasoning with deterministic real-world capabilities (PDF, Vision, Speech, Web Search, Shell, Computer Use)
- Skills — reusable instructions in the SKILL.md format (Agent Skills / PPM), loaded on demand by the model
- A2A — expose your agent graphs to other agents and call remote ones (Agent-to-Agent protocol 1.0)
- Production controls — tool-call guardrails, evals, OpenTelemetry tracing, persistent memory, and calibrated decisions with Jev
- FMX Visual Components — drop-in UI for multimodal chat interfaces
- Universal Connector — switch providers at runtime without changing your application code
Whether you need a simple one-provider integration or a multi-agent, multi-provider, retrieval-augmented production system, MakerAI covers the full stack — natively in Delphi.
🚀 What's New in v3.9
OpenAI GPT-Live — full-duplex voice with delegation
TAiOpenAiLiveChat (gpt-live-1, DriverName OpenAiLive) listens while it speaks: you can interrupt it, and it hands the heavy thinking to another model — a Responses model run by OpenAI (web search, your TAiFunctions and MCP tools executed locally) or any TAiChatConnection through DelegateChat (Claude, Ollama, an agent graph). Conversation turns are rebuilt from the session timeline, so overlapping speech no longer splits answers. Demo 092-GPTLiveVoice: microphone, speaker, local functions, RAG over MCP with a read-only guardrail, anti-echo speaker mode and a quick guide. Tested live.
Jev decisions on local models (Ollama)
Ollama 0.35+ serves the Jev contract at /v1/systemone with local decision models — nimble, tev1 and Cloudflare's clef/clef-flash. TAiJev only needs Url := 'http://localhost:11434/v1/'; the eight Jev adapters gained Url, TAiJev.Ask takes images for Clef (PNG/JPEG/WebP), and the TypeSafe price is no longer charged on other servers. Tested with nimble (routing, prompt guard, batch labeling) and clef 27B on images.
Memory leaks fixed
An MCP server leaked every tool schema on each tools/list (it grew forever in production), the MCP client leaked the tools/list response on each Initialize, and TAiRealtimeFactory never freed its dictionary.
⚠️ Behaviour changes
- OpenAI audio defaults:
TAiOpenAiAudio/TAiOpenAiSpeechToolusegpt-transcribeandgpt-4o-mini-tts(OpenAI shuts downwhisper-1on 2027-02-26 andtts-1on 2027-01-06).gpt-transcribereturns JSON only — srt/vtt/timestamps are reported inTTranscriptionResult.Warning; settmWhisper1to keep them until the shutdown. - Custom MCP tools:
IAiMCPTool.GetInputSchemareturns a new object owned by the caller (what the built-in tools already did). A custom tool that returned a cached field must return a copy. - OpenAI Responses driver:
Max_Tokensis now sent asmax_output_tokens(it was silently ignored), so a low limit can cut answers that used to be complete.
What came in v3.8
Released 2026-09-29. Six items change existing behaviour; they are marked ⚠️ below — two of them
(TLS certificate checks on POSIX and the IAiMemoryStorage signature) can require code changes.
Jev — calibrated decisions before spending an LLM
TAiJev (Source/Tools/uMakerAi.Jev.pas) wraps Jev, TypeSafe AI's "System One" model.
Jev does not generate text: it answers typed questions — Choice, Score, Noul (yes/no) —
with calibrated probabilities and a confidence value your code can threshold. It is meant for
the fast, cheap decisions that today cost a full LLM round-trip: routing a query to the right
specialised agent, deciding whether that agent needs its RAG variant, classifying, gating.
Several questions travel in one request (~700 input tokens at US$0.042 per million). Questions
are validated locally, 429/529 are retried with backoff, and the model is pinned to
jev-1.13.0 so thresholds stay valid. Demo: 084-JevRouter.
For agent graphs, TAiJevRouterTool (Source/Agents/uMakerAi.Agents.Tools.JevRouter.pas) is a
node tool that writes the chosen route to Blackboard['next_route'], so an existing
lmConditional link follows it — no engine change. Low confidence or a failing API falls back
to NextNo instead of breaking the graph, and yes/no flags asked in the same call land in the
blackboard for lmExpression or an OnRoute handler. Twenty-one new regression cases run all
of it offline against a fake transport.
Two more hand-offs make Jev a drop-in for decisions the framework already takes:
- SmartDispatch without the classification LLM call.
ChatTools.DispatchClassifier(new,
provider-neutral) receives the tags whose tools are assigned;TAiJevDispatchClassifieranswers
with Jev. Tool requests skip one LLM round-trip, and CHAT replies are now generated with the
conversation history (the LLM pass answered in an isolated two-message context). Unsure or
failing → the usual LLM pass. 15/15 on Spanish and English requests. -
Semantic guardrails.
TAiGuardrails.Classifier(new) judges the tool calls the allow/block
lists let through;TAiJevGuardrailClassifierblocks when P(risk) ≥ 0.5. Lists still catch the
enumerable (rm -rf) for free; Jev catches what no list anticipates — an e-mail carrying a
password to an outside address, a transfer to an unknown account. Safe calls scored ≤ 0.17 and
harmful ones ≥ 0.88 across 13 calibration cases. Fails closed by default. Optional permission
categories (read,write,financial,system, …) are judged in the same call: block whole
categories, audit each call viaOnCategorized, and describe domain-named tools with
ToolDescriptions(20/20 on 20 calibration calls). -
Input guardrail.
ChatTools.PromptGuard(new) checks the user's message before it reaches
the LLM, right after the existing regex sanitizer.TAiJevPromptGuardasks in one call about
prompt injection, credentials in the message, harmful requests and — given aScope— off-topic
questions. On nine test messages the regex caught 1 of 6 problematic ones; Jev caught all 6 and
let the greeting and the legitimate questions through. A blocked message never touches the network.
Demo: 085-JevDispatchGuard.
And two more for quality and retrieval:
- Calibrated eval judge.
TAiEvalRunner.Scorer(new) answersExpectScore('criterion', 0.7)
with the probability that the output meets the criterion — faster and cheaper than an LLM
judge, and the bar is set in code.TAiJevEvalScorer: passing answers ≥ 0.97, failing ≤ 0.02
on 10 calibration pairs. - Semantic reranking for RAG.
TAiRAGVector.Reranker(new) replaces the cosine second stage
of VQLRERANK:TAiJevRAGRerankerscores each passage for usable evidence and drops
passages that try to instruct the model (prompt injection). No embeddings are recomputed;
if the reranker fails, the search falls back to cosine. Demo:086-JevEvalsRag. - Bulk labeling.
TAiJevBatchLabelerruns the same questions over many rows in parallel and
returns labels, confidence, top-3 suggestions, total cost and the rows worth a human look; a
failing row never stops the batch. Validated by reproducing the prototype: 126 accounting
entries against 225 chart-of-accounts codes in 4.2 s for US$0.05 — 92.9% overall and 100% on
the 61% it would book automatically (confidence ≥ 0.8). Demo:087-JevBatchLabeling. - Model routing.
TAiJevModelRoutersends each request to the cheapest model that can handle
it. Jev describes the request (task, difficulty, sensitivity); readable rules in code pick the
tier. Switching provider on aTAiChatConnectionused to drop the conversation — the router
migrates the text history, so a chat can start on Groq and escalate to Claude and still
remember the first turn. Live, with Groq → DeepSeek → Claude Sonnet → Opus: 7 of 8 answers
judged useful, the trivial ones in under a second. Demo:088-JevModelRouter. - Metering for usage-based billing.
TAiJevand all eight adapters exposeUsage(requests,
input/output tokens, cost in USD) and anOnUsageevent fired once per operation — one guard
check, one rerank, one batch — on the thread that ran it, so a server can charge the right
customer. The parallel reranker reports a single total after joining its calls. Previously only
the batch labeler exposed tokens; the other adapters received them and dropped them.
Also fixed:
lmExpression parsed numbers with the regional settings only, so on a Windows
using a decimal comma '10.25' > 9.5 was compared as text and returned False. It now falls
back to a decimal point.
Computer Use on Linux
TAiLinuxExecutor (Source/Tools/uMakerAi.Tools.ComputerUse.Linux.pas) drives X11 through
xdotool and captures with scrot, covering the 19 canonical actions with the same
public interface as the Windows and macOS executors. The framework itself needed no change:
TAiComputerUseTool only ever used the RTL and delegates everything to
OnExecuteAction / OnRequestScreenshot, so it cross-compiled to Linux64 untouched — what
was missing was only the executor. Runtime-tested on Xvfb with gpt-6-astra and
claude-opus-4-8, driving both a text editor and Chrome. Demo: 083-ComputerUseLinux.
Computer Use delegation — handing the call to a remote client
A headless process (a broker on a VPS) can now claim a computer_call and forward it to
whoever actually owns a screen. The contract already existed in the base class and in Claude
— fill ToolCall.Response from OnCallToolFunction and the driver does not execute
locally — but OpenAI ignored it and Gemini never even fired the event. Both now honour
it. For OpenAI the delegation is atomic over the batch: gpt-6-astra sends an array of
actions that admits a single computer_call_output, so splitting it between a local and a
remote executor would leave the final screenshot ownerless. Demo: 082-ComputerUsePassthru.
⚠️ TLS certificates are now verified on POSIX
TOpenSSLTransport used to run with SSL_VERIFY_NONE — it accepted any certificate,
from anyone, and that is the transport the Realtime module uses on Linux and macOS. It now
verifies the chain against the system CA store and the hostname (SSL_set1_host;
SSL_VERIFY_PEER alone validates the chain but not that the certificate was issued for the
host you dialed). Verified against badssl.com, 7/7, including the wrong.host case.
> Breaking: pointing Realtime at an endpoint with a self-signed certificate (a local LM
> Studio, an internal proxy) now fails until you set InsecureSkipVerify := True.
⚠️ Gemini executed no user functions at all
TAiGeminiChat.DoCallFunction had its inherited commented out and answered
'Command not found' to every non-Computer-Use tool call — and it is the driver's
only dispatch point. Neither AiFunctions nor OnCallToolFunction ever ran. Restored.
> Behaviour change: tools that silently returned "not found" with Gemini now execute.
⚠️ TAiShell crashed on non-English Windows
Shell output was decoded with TEncoding.UTF8.GetString, which validates its input and
raises EEncodingError. cmd.exe writes in the console OEM codepage (cp850 on a Spanish
Windows) and even its own banner carries accents, so the component died on the first
command. Invisible in English (pure ASCII) and on Linux (bash does emit UTF-8). Decoding
now falls back to the OEM codepage. Two more, found while exercising it on Linux: stderr was
dropped whenever the sentinel arrived in the same read, and a timeout left the session
permanently unusable (the Restart was written but commented out).
⚠️ TAiMemory: namespaces are now enforced on id-based operations
TAiMemory isolated namespaces in its searches, but Get, Update, Delete, Link and
Unlink addressed memories by id alone — and ids are sequential, so an agent could read,
change or delete another namespace's memories by guessing one, including through the
memory_delete / memory_link MCP tools. ImportFromJSON also honoured the namespace field
of the JSON and could write into another namespace. Every id-based operation now requires the
active namespace (a foreign id behaves exactly like a missing one), and imports always land in
the active namespace. Reported responsibly in #127.
> Breaking for custom storages: the id-based methods of IAiMemoryStorage now take the
> namespace. The bundled SQLite storage is updated; a custom implementation must add the
> parameter.
⚠️ A forced tool_choice applies to the first call only
With Tool_choice := 'required' (or a named function) the driver re-sent the forcing on
every round, including the one that returns tool results, so the model had to call a tool
again each time and the agentic loop never ended — 231 requests on claude-sonnet-5 before
it was killed. TAiChat.Tool_choice now reads auto in those rounds, which fixes every driver
at once. The first call of a turn is still forced.
⚠️ TLLMNode.DriverName now defaults to empty
The constructor used to set DriverName := 'Claude', which always overrode the driver of an
assigned TAiSkill. It is now empty and resolves node → skill → 'Claude', so a node without
a skill behaves exactly as before and forms saved with the old default keep 'Claude'
explicitly. See the skills section below for the rest of the precedence fix.
Qwen — native Alibaba Model Studio driver
TAiQwenChat (Source/Chat/uMakerAi.Chat.Qwen.pas, driver name Qwen) talks to DashScope's
OpenAI-compatible endpoint (international region by default; key in DASHSCOPE_API_KEY).
Hybrid Qwen models think by default on the API side, so the driver always sends
enable_thinking: off unless cap_Reasoning is in ModelCaps, with ThinkingLevel mapped to
thinking_budget. Thinking-only models (qwq-plus, *-thinking-*) and open-weight models
(which only think when streaming) are handled for you. Registered models are the ones tested
live: qwen3.8-flash (default), qwen3.8-max, qwen3.7-plus, qwen3-vl-flash,
qwen3.8-omni-flash with vision; qwen3-max, qwen-plus/flash/turbo, qwen3-coder-plus/flash;
and qwq-plus, which only answers when streaming and is registered as asynchronous.
Beyond chat, the same key covers the rest of Model Studio, all tested live from Delphi:
- Image generation and editing (
cap_GenImage):qwen-image-3.0,qwen-image-2.0,
qwen-image-max,z-image-turbo,wan2.7-image. Attach one to three images to the prompt and
the same call edits them (qwen-image-edit-plusby default), keeping the original aspect ratio
unless you set a size. The driver downloads the result intoMediaFiles. - Video (
cap_GenVideo, Wan): text to video, image to video (attach one image) or first
and last frame (attach two). 720p by default instead of the API's 1080p;VideoParams.Params
passes through (duration, audio, seed...). Wan 2.5+ videos come with an audio track. - Text to speech (
cap_GenAudio):qwen3-tts-flash, voice and language fromTtsParams. - Transcription:
qwen3-asr-flashincmTranscription, where the prompt is passed as
context so proper names come out right. A text model withcap_AudioinSessionCapsuses it
to transcribe before answering, andqwen3.8-omni-flashunderstands audio natively. -
Realtime (
TAiRealtimeConnection, new unituMakerAi.Realtime.Qwen):DriverName := 'Qwen'
for voice conversation,'QwenSTT'for live transcription and'QwenTranslate'for simultaneous
realtime module changed. Driver-specific settings (voice, instructions, target language) go in
the connector's newDriverParams, which works for every realtime driver.TAiQwenRealtimeTTSgoes the other way — text in, audio out while it
is generated — so an LLM answer can be spoken as it streams (first audio ~0.5 s after the first
text; 2.5 s from question to voice with qwen3.8-flash).
TranslateFrom,TranslateDomain, aTranslateTermsglossary) and each message is translated.
The driver sends only the latest message, as the API requires, and fixes plus/turbo streaming,
which repeats the accumulated text in every chunk.
Demo089-QwenShowcasewalks through all of it with one key. -
Custom voices:
TAiQwenVoicesclones a voice from a sample or designs one from a text
description, and lists or deletes them. Put the returned id inTtsParams.Voice; the driver
switches to the TTS model that voice requires. Clone only voices you have consent to use. - Embeddings:
TAiQwenEmbeddings(driverQwen),text-embedding-v4by default. - Rerank:
TAiQwenRAGRerankerplugsqwen3-rerankintoTAiRAGVector.Reranker, up to 500
passages per call, with an optional task instruction.
Skills — SKILL.md, loaded on demand
MakerAI now speaks SKILL.md, the format of Claude's Agent Skills and of the PPM registry's
skill packages: YAML frontmatter (name, when to use it) plus Markdown instructions. One parser
(uMakerAi.Skills.Format) serves the whole framework, from a string, a file, a folder or the
registry.
TAiSkills(new,Source/Tools/uMakerAi.Tools.Skills.pas) gives any chat with function
calling on-demand skills. The model sees only a one-line-per-skill catalog in the description
ofuse_skilland loads the full instructions when it needs them, so dozens of skills cost no
tokens until used. Folder-based skills can ship supporting files, read withread_skill_file,
which is confined to the skill folder (no.., absolute paths, links or binaries, never
executes anything). Veto and audit events; guardrails apply as to any tool. Live, 9/9 on
OpenAI, Claude and Groq: the right skill for each request, none for an unrelated one, and
use_skill→read_skill_filechained when the instructions asked for it.TAiSkill.FromPPMworks. It requested a URL that returns the website's HTML and a JSON
format no package publishes; it now downloads the real SKILL.md ('code-review'also finds
skill-code-review). NewFromFolder/FromSkillFile; the JSON format still works for local
files. A SKILL.md never supplies an API key.TLLMNoderespects its skill. NewResolveConfig: driver from the node, else the skill,
else Claude; the skill's model only if it belongs to that driver; the skill's system prompt
followed by the node's. Previously the skill's driver was always overridden and switching
driver lost the skill's model and key.TAiPromptsloads local skills (LoadSkillFromFile,LoadSkillsFromFolder) and
ApplySkillcopies one into anySystemPrompt. Registry versions are now picked by semver.
Guides: Docs/Version 3/uMakerAi-Skills.EN.md (Spanish: uMakerAi-Skills.md). Demo 091-Skills.
Also
- Every driver on the shared streaming parser returned the answer twice. In asynchronous
modeOnReceiveDataEndreceived'Done'#13#10'Done': the end of the stream re-added the
accumulated text. Affected all drivers built on the common parser (verified on Groq and
DeepSeek; Kimi, Grok and Mistral now return it once too). - Files generated by Groq's code interpreter were lost when streaming.
gpt-osssends
executed_toolsinside each delta, in two chunks per tool; the parser only looked at the
top level, the format of the retiredgroq/compound. They are now merged and delivered as
in the synchronous path. - Groq catalog brought up to date with the API.
llama-3.1-8b-instantand
llama-3.3-70b-versatileno longer exist and were the driver's default, so a Groq connection
without an explicit model came back empty. The default is nowopenai/gpt-oss-20b, and both
old names are aliases (togpt-oss-20bandgpt-oss-120b), so existing code keeps working.
groq/compoundwas removed without a replacement.qwen/qwen3.6-27bwas registered with a
token limit the API rejects; fixed, andqwen/qwen3.8-27badded. Demo 014 now runs Groq's
code interpreter ongpt-oss-20b(--groq). - Claude sometimes sends coordinates as a JSON array and sometimes as a string containing
one — within the same turn.TryGetValuemissed the second form, the
coordinate was lost and the action landed on (0,0): a click in the screen corner, after
which the model retried until it ran out of turns. Affectedcoordinate,
start_coordinateandregion, i.e. click, double/triple click, drag and zoom. - OpenTelemetry validated against a real collector for the first time (Jaeger): the trace
crosses the A2A boundary, sotraceparentpropagation through_metaworks. Fixed
otel.scope.version, hardcoded to3.5while the framework was on 3.7. - Demos 031 and 077 now build and run on Linux64 — which also makes FireDAC on Linux a
tested path (PostgreSQL 18 + pgvector 0.8.1 throughlibpq.so.5). - Documented that the Realtime module needs
CheckSynchronizein console apps and services:
every event is dispatched withTThread.Queue, so without a message loop no event ever
fires — not evenOnError.
What's New in v3.7
Computer Use, Refreshed on Both Live Providers
Both vendor APIs changed under our feet, and one of them had gone silently dead. Claude was broken: the driver still declared computer_20251124, a tool type the Anthropic API now rejects for every model. It is now computer_toolset_20260801, which takes no parameters and explodes the old single computer tool into 17 individually named tools. OpenAI joins natively with gpt-6-astra and the parameterless computer tool, which sends a whole batch of actions per turn — the driver runs them in order and answers with one final screenshot.
Both APIs converged on the same design: the tool declares no screen dimensions (the model infers them from the screenshot) and coordinates come back as pixels of the image you sent.
GLM (Zhipu AI / Z.ai) — New Provider
TAiGLMChat brings a 14th provider: OpenAI-compatible endpoint, explicit thinking control (the API ships it ON by default, the driver decides), reasoning_content captured and re-sent across turns, and free tiers (glm-4.7-flash, glm-4.6v-flash).
Correctness Pass
RAG on pgvector (OFFSET was trimming instead of paginating; the vector did not survive the round trip), token and cache accounting across OpenAI, Claude and MakerAi streams, a race where the model travelled through the global registry, tool-calling continuation off the main thread, and reasoning no longer leaking into the user-visible answer.
What's New in v3.6
MCP Specification 2026-07-28 — Stateless, Dual-Era
The Model Context Protocol dropped sessions and the initialize handshake. MakerAI implements the new stateless revision on both sides and keeps talking to legacy peers: clients probe with server/discover and fall back automatically; the server serves modern per-request _meta requests statelessly while the legacy handshake and session gating keep working. Includes the MRTR pattern, so a tool can pause and ask the user for confirmation (OnInputRequired on the client, TAiAuthContext.InputResponses on the server).
Observability — OpenTelemetry Tracing
TAiTelemetry exports OTLP traces to any standard collector (Jaeger, Grafana Tempo, Langfuse, Arize Phoenix) following the GenAI semantic conventions. Spans cover chat turns with token usage, tool executions, agent graphs and nodes, RAG retrieval and MCP requests — with W3C traceparent propagated through MCP _meta, so a client and a server in different processes share one distributed trace. Opt-in, zero overhead when disabled.
A2A — Agent-to-Agent Protocol (first Delphi implementation)
If MCP is the agent-to-tool layer, A2A (Linux Foundation) is the agent-to-agent layer. TAiA2AServer publishes any agent graph as a standard A2A agent (Agent Card + JSON-RPC), TAiA2AClient consumes remote agents, and TAiA2ARemoteAgentTool
federates: a node in your graph can delegate its work to a remote agent — including one written in another language or framework. Demo: 072-A2AFederation.
Guardrails & Evals
TAiGuardrails intercepts every tool call before it executes (allowlists, blocklists, forbidden argument patterns, programmatic veto) — blocked calls never run and the LLM gets the reason so it can replan. TAiEvalRunner brings systematic evaluation: fluent test cases against any target, deterministic checks plus optional LLM-as-judge, with ToJSON reports for CI.
First Automated Regression Suite
Tests/RegressionSuite/ — 17 in-process cases covering MCP, agents, A2A, guardrails and evals. No API keys, under a second, exit code for CI. Built on TAiEvalRunner itself.
What's New in v3.5
Typed ModelConfig Channel
Capability configuration now lives in a single typed surface: ModelConfig.ModelCaps / SessionCaps / Tool_Active / ThinkingLevel moved out of the string-based Params/RTTI channel, with per-field user pins and transparent compatibility migration — existing code keeps working unchanged.
Full-Duplex Voice Suite
TAiGrokRealtimeChat— xAI Grok Voice speech-to-speech (function calling, session resumption with replay, binary audio transport, ephemeral tokens)TAiRealtimeVoiceBase— shared full-duplex base; voice events flow through the universalTAiRealtimeConnection- gpt-transcribe / gpt-live-transcribe — OpenAI's Whisper successors, fully integrated
August 2026 Provider Refresh — All 9 Cloud Providers, Runtime-Tested
Claude 5 family (adaptive thinking, FastMode, compaction, server-side fallbacks) · Gemini 3.5/3.6 + Nano Banana GA · Mistral Voxtral TTS + OCR 4 · Kimi K3 · DeepSeek V4 (explicit thinking control) · Cohere Command A+ · Groq qwen3.6 · xAI grok-4.3/4.5/build — with retired-model cleanup and compatibility aliases throughout.
Grok Native Video & Image Generation
TAiGrokChat now generates video with grok-imagine (async job + polling + mp4 as TAiMediaFile, new VideoDurationSeconds property) and images with grok-imagine-image — activated by cmVideoGeneration/cmImageGeneration or the [cap_GenVideo]/[cap_GenImage] gaps.
What's New in v3.4
Delphi 13.1 Florence Support
v3.4 is fully tested and compatible with Delphi 13.1 Florence (CompilerVersion 37.1), in addition to the existing range from Delphi 10.4 Sydney through Delphi 13 Florence.
Selective Driver Registration
The biggest infrastructure change in v3.4: TAiChatConnection no longer force-loads all providers at startup. Each driver now self-registers only when explicitly imported, eliminating unnecessary initialization overhead:```pascal
// Load only what you need
uses uMakerAi.Chat.AiConnection, uMakerAi.Chat.OpenAi, uMakerAi.Chat.Claude;
// Load all drivers at once (legacy behavior)
uses uMakerAi.Chat.Initializations;
```
Real-Time STT — TAiRealtimeConnection
New universal connector for real-time speech-to-text via WebSocket:
TAiRealtimeConnection— provider-agnostic STT connector; switch providers viaDriverNameTAiOpenAiRealtimeSTT— full OpenAI Realtime API implementation (24 kHz PCM16, VAD modes, streaming transcription)- Pure-Pascal WebSocket client with native TLS via Windows SChannel — no extra DLLs required
- Thread-safe PCM16 resampler; supports push-based audio streaming from any source
GPT-Transcribe — Next-Gen OpenAI Transcription (Whisper successors) 🆕
OpenAI's new transcription models (Aug 2026) are fully integrated — better accuracy on real-world audio, accents, numbers, specialized terminology and loud background noise:
| Model | Use case | Word Error Rate |
|---|---|---|
gpt-live-transcribe |
Live low-latency STT (Realtime WebSocket) | 9.60% (vs 11.65% Whisper) |
gpt-transcribe |
Completed files and batch workloads | 8.98% (vs 15.21% Whisper) |
TAiOpenAiAudiogainstmGptTranscribe/tmGptLiveTranscribewithTranscriptionKeywords+TranscriptionLanguagesfor REST/batch transcription- Legacy models (
whisper-1,gpt-4o-transcribe) remain available — they're still required for subtitles (SRT/VTT), word timestamps and diarization (gpt-4o-transcribe-diarize), which the new models don't support; the components degrade formats safely per model - VoiceBridge demos (062–065) migrated: live channels use
gpt-live-transcribewith contextual prompts and guided language detection; diarized channels stay ongpt-4o-transcribe-diarize
Grok Voice — Real-Time Speech-to-Speech (xAI) 🆕
…(正文过长,页面只展示前面部分,完整内容见仓库)
