Thunderbird answers from bounded mailbox context, not hand-wavy chat memory. It parses, verifies sender auth, scans attachments, redacts, ranks, cites, and only then sends the final Assistant request to a configured Ollama endpoint.
This is not a generic chatbot bolted onto a mailbox. The product is designed around grounded retrieval, visible processing stages, bounded context, and security evidence that users can inspect.
Ask about selected messages, a folder, an account, or all analyzed mail. Thunderbird narrows scope, composes evidence, and keeps citations attached to the answer path instead of turning the inbox into a detached chatbot transcript.
Retrieval ranking, compression, and context budgeting happen before inference, so the model sees the right slices of mail instead of a reckless full-folder paste.
DKIM, SPF, DMARC, sender trust, URL findings, and bundled ClamAV attachment scan results can feed a bounded security view, helping users understand why a message looks safe, suspicious, or incomplete.
FLAN-T5 handles local summaries, MiniLM supports embeddings and RAG, GLiNER covers PII, and the native classifier keeps fast local categorization available even before Assistant synthesis.
Users can inspect endpoint health, active work, backlogs, trace payloads, and runtime logs instead of treating AI as a black box.
Generated files, traces, caches, and model artifacts stay visible from the product, with deletion and reset flows that do not require spelunking through hidden directories.
The landing page now shows real captures from the app: the Assistant runtime view and the specialist local roles that back summaries, privacy, and retrieval.
Scope, indexed RAG mail, attachment-scan backlog, and endpoint probe state are all visible from the Assistant surface instead of hidden behind background jobs.
FLAN-T5 stays visible as a dedicated local summary role.
GLiNER keeps privacy detection bounded to a named role.
MiniLM powers embeddings and retrieval-oriented indexing.
Before any answer is synthesized, Thunderbird can already parse the message, inspect trust and sender signals, verify sender auth, evaluate URLs or attachments, redact PII when required, and attach bounded evidence to the final decision path.
Mailbox selection, parsing, security evidence, template extraction, retrieval planning, citations, redaction policy, and prompt assembly remain inside Thunderbird.
The configured Ollama endpoint receives the final bounded request, not the entire mailbox, not hidden product state, and not an unreviewed flood of raw content.
Runtime traces expose endpoint URL, model, retry state, and retrieval evidence. Data Governance keeps generated artifacts visible and removable.
DKIM verification, SPF header evidence, and DMARC policy plus alignment checks are part of the mail-native security story, including alignment against the RFC5322 From domain.
Bundled ClamAV attachment scanning runs as part of Thunderbird’s own security path, with asynchronous queueing so mail review does not block on a single large attachment.
GLiNER and deterministic rules keep sensitive content reviewable, redactable, and visible as a separate mailbox role instead of burying privacy decisions inside a generic prompt.
FLAN-T5 summaries, MiniLM embeddings, and the Ollama-backed Assistant work as one layered flow: summarize locally, index for retrieval, then ask questions over bounded evidence.
The model is the last stage, not the whole product. This keeps AI grounded, keeps security evidence available even during failure, and makes debugging practical when an endpoint or local stage degrades.
Selected messages, folder, account, or all analyzed mail.
Canonical body, MIME structure, trust, and sender context.
DKIM, SPF, DMARC, AV, URL, and trust evidence is attached.
Templates, regex hits, categories, PII, summaries, and embeddings.
Rank relevant messages and build a bounded evidence pack.
Send the final request to the configured Ollama endpoint.
The product promise is simple: if AI touches your mail, you should be able to see what it saw, what it sent, what it stored, and how to remove it.