Status and roadmap
HelpChatAI is work in progress, developed in the open and already usable: the engine is complete and covered by tests, and the components work in real applications. This page says plainly what is done, what is verified, and what is not there yet — so you can decide whether it fits your project today.
What is done
| Area | State |
|---|---|
| Engine (store, worker pool, cancellation, persistence) | Complete, covered by tests |
| Providers | stub, docsearch, claude, openai (and with it Azure, Groq, OpenRouter, Mistral, Ollama...), gemini; the three model providers with streaming, grounding and tool use |
| Documentation index and builder | Complete; tested on a real 38-page Italian manual. Areas of the manual, from the sub-folder of a page or its front matter, for an assistant that answers from one part of it |
TVclHelpChatAI | Complete: bubble, F1, help viewer, context, embedding, VCL styles, position memory |
| Documentation pages in the panel | Complete: a folder of Markdown pages rendered in the chat, with OnOpenLink for everything that leaves it |
| Dictation | Complete on VCL: the dictation of Windows, or the speech recognizer where it still transcribes |
TFmxHelpChatAI | Complete on Windows; the other platforms compile but are untested |
| Local transport (127.0.0.1) | Complete |
| HelpChatAI server and remote transport | Complete: sessions per user, memory or disk storage, application keys |
| Actions | Complete, with a local engine and with a remote one, on claude, openai and gemini: the actions of one assistant, the ones every assistant shares, and AllowedActions in the configuration to narrow both |
| Skills | Complete, with a remote engine too, VCL and FireMonkey (LiveBindings included): forms, commands, records, pages, named in Settings.Skills |
| What an assistant knows | Complete: Expertise with its rule, Instructions, every text of the prompt a property, OnBuildSystemPrompt, HelpChatSetMsg |
| Prompt caching | Complete on claude, openai and gemini: a system prompt that does not change, the context with the question, the cache breakpoints of Claude, the tokens of every request through HelpChatOnTokenUsage |
| Attachments | Complete with the local engine: PDF, images and text files, on claude, openai and gemini |
| Saving and copying | Complete on Windows: an answer, the conversation or a documentation page copied with its formatting, or saved as a PDF with a header and a footer of yours. The PDF needs a client that can write a file, so not with a remote engine |
| Packages for Delphi 10.4, 11, 12, 13 | Built with each compiler |
| Installer (InnoSetup) | Complete |
| Design time | Splash and About box, component editor with the configuration editor (six pages), live preview in VCL and FireMonkey, configuration check, property editors |
| Several assistants | Complete with the local engine: one component per form, a Profile each in one configuration file, one engine per profile, F1 to the assistant of the active form |
| Models | Complete: a list in the configuration, the switch from the header of the panel, the icon of each model, the choice remembered per user and per profile |
| Languages | Complete: six languages, chosen from the flag in the header of the panel, and a manual in another language than the chat |
| Logging | Complete: Settings.LogLevel (none, info, detail, debug), OnLogEvent on both components, HelpChatOnLog for code with no component; the demos write it to a file |
| Slash commands | Complete: nineteen commands, EnabledCommands, and every one of them from code |
| Sample applications | The Windows 10 CRM of Embarcadero, in both frameworks, sharing one data module and one manual, each with a multi-assistant edition |
616 automated tests run without network on Win32 and Win64: the store, the runner (streaming, cancellation, shutdown with a stuck provider), the index scoring, the Markdown safety, the SSE parsing and request building of each of the three model providers, the documentation pages, the configuration check, both transports over real HTTP, the actions from a real worker thread, the skills on forms built in code, a whole conversation replayed from the streams of the real service, the secrets of the settings and the printable document - two of which write a real PDF with the WebView2 runtime of the machine, and skip themselves where there is none.
What is not there yet
- A file offered by the application through the chat. The assistant can now write a document and hand it over as a PDF, but a file produced by the application itself - a report, an export - still cannot be offered inside the conversation.
- The PDF with a remote engine. The document would be built on the server, where there is no screen to save it from: the buttons say so (HTTP 501) instead of doing nothing. Copying works there.
- Attachments with a remote engine, which has not been tried: the page talks to the server directly and the documents would live in its session.
- Models and languages with a remote engine. The engine of a profile is shared by every user of that assistant, so neither the model nor the language can be switched from a client. The profiles themselves work: one engine per profile of the server's configuration (assistants on a server).
- FMX on macOS, iOS and Android. The code compiles and uses the native web view of each platform, but only Windows has been tried. On Windows FireMonkey is now exercised by a sample application of its own, which is how two defects of that half were found.
- A retriever based on embeddings. The current one is deterministic keyword scoring behind
IHelpChatRetriever; a vector-based one can replace it without touching the providers. - Providers beyond the three APIs that ship. Anthropic, Google and everything that speaks the OpenAI API are covered; anything else is a unit plus a
RegisterProvidercall - the provider contract is small and public. - A dictation engine of our own. Windows offers two and neither is ideal: its own dictation types into the control and tells the application nothing, while
Windows.Media.SpeechRecognitionreports the text but on recent Windows 11 builds can hear a voice and transcribe nothing. The engine that drives Win+H is internal and has no public metadata, so it cannot be called. A cloud service or a local model would be a thirdTHelpChatVoiceEngine, behind the same interface. - Dictation on FMX. The microphone button is VCL only for now.
Milestones so far
| M1 | Engine, providers, tests, a console harness |
| M2 | Field test with the real Claude provider: streaming, grounding, screen context |
| M2b | Index builder: keywords, stemming, idf ranking, excerpts; tested on a real manual |
| M3 | Transport abstraction, local server, HTML/htmx panel, TVclHelpChatAI, bubble, F1, design-time package |
| M3b | VCL style theming, F1 through the Delphi help system, position memory, the chat as a control embeddable in any container |
| M4a | TFmxHelpChatAI, host commands through the server |
| M4b | HelpChatAI server (sessions per user, memory or disk), remote transport chosen in the configuration file, loopback host channel |
| M5 | Packages for Delphi 12, 11 and 10.4, InnoSetup installer |
| M6 | Actions: registered by the application, Anthropic tool use, execution on the application's thread with confirmation |
| M7 | Question dictation, the library in the splash and About box of the IDE, component editor with live preview and configuration check, property editors |
| M8 | The openai and gemini providers, streaming and tool use like claude |
| M9 | Caption and AssistantName, the assistant icon in the header of the panel, the FireMonkey bubble round and with its icon, and the CRM sample application in both frameworks |
| M10 | The FireMonkey demo complete: lead, proposal and draft windows, all three actions, F1 from every window |
| M11 | Attachments: the user hands the assistant a PDF, an image or a text file, and an action turns what it read into a record |
| M12 | An answer that leaves the panel: the clipboard with its formatting, a PDF of an answer, of the conversation or of a page of the manual, with a header and a footer of the application |
| M13 | Several models to switch between from the panel, each with its icon; slash commands in the question box, and everything the panel does available from code |
| M14 | Six languages, chosen by the user in the panel; a manual in another language than the chat |
| M15 | Several assistants in one application: profiles in one configuration file, one engine per profile, areas of the manual, shared and allowed actions, two nodes for every secret, the configuration editor |
| M16 | An expert, not only a manual: Expertise and its rule, the name and the time in the prompt, every text of the prompt a property, the multi-assistant edition of the demos, the preview of the FireMonkey component |
| M17 | Skills: abilities written once by the library and named by an assistant; the bubble of the main form in front of every form |
| M18 | Prompt caching on the three model providers, measured: the system prompt the same for every question, the context with the question |
Contributing
The repository is EtheaDev/Delphi-HelpChatAI. Issues and pull requests are welcome — in particular:
- trying FMX on macOS, iOS or Android and telling us what breaks;
- providers for other models (the three that ship cover Anthropic, Google and everything that speaks the OpenAI API, but not, say, a model behind your own gateway);
- translations of the interface strings into more languages (today English, Italian, Spanish, French, German and Portuguese, in
HelpChatAI.Messages), and native speakers reviewing the four added last; - real-world feedback on what the retriever finds for your manual, which is the part that decides whether the answers are good.