docs

Documentazione tecnica

Come è fatta Déjà sotto il cofano. Estratto dal vault di progetto.

#Architettura

main.py avvia QApplication e 3 thread daemon coordinati da un unico stop_event:

t_captureloop screenshot + OCR ogni 5s
t_indexerembedding pendenti → tabelle vec ogni 10s
t_audiorecord mic+loopback, segmenti 30s, transcribe

La GUI Qt sta sul main thread. L'hotkey gira su un thread separato ed emette un signal che fa il toggle dell'overlay.

#Pipeline di cattura

capture.py

1shot = mss.grab(monitor)        # multi-monitor aware
2jpeg = compress(shot, q=70)     # resize 75%
3if sha256(jpeg) == last: skip   # dedup
4text = tesseract(shot, "ita+eng")
5text = redact_pii(text)         # cards, IBAN, tax ID, email, phone
6db.insert(screenshots, ts, app, jpeg, text)

#Ricerca semantica + esatta

search.py

1query text
2  → encode(normalize=True)                 # 768-dim
3  → sqlite-vec KNN:
4      WHERE emb MATCH vec_int8(?) ORDER BY distance LIMIT k
5  → score = 1.0 - distance                 # cosine
6  → thresholds: screenshot ≥ 0.18, audio ≥ 0.15
7  ‖ PARALLEL: tokenized exact match (LIKE AND, fuzzy)
8  → merge dedup: exact first → score desc → ts desc

#AI & RAG

Endpoint OpenAI-compatibile. L'assistente decide quando interrogare la memoria via tool calling, poi cita i ricordi con marker [ss:ID] / [au:ID] resi come card inline.

ai_assistant.py

1def chat_stream(history, user_msg):
2    """yields (kind, content): text | tool_call | tool_result | error"""
3    # max 5 tool-calling iterations
4    # tools: search_memories, list_recent, list_by_date_range, memory_stats
5    # auto-trim history if estimated tokens > context budget
6    ...

#Stack tecnico

GUI	PyQt6 — QSystemTrayIcon, QStackedWidget
DB	SQLite WAL + sqlite-vec (int8 cosine, 768-dim)
Embeddings	paraphrase-multilingual-mpnet-base-v2 (768-dim)
STT	faster-whisper large-v3-turbo (CPU int8)
OCR	Tesseract (ita+eng)
Audio	pyaudiowpatch (loopback, callback-based)
AI API	OpenAI SDK — Qwen3-235B / Kimi-K2.6, 128K ctx
Vision	Gemini 2.5 Flash (OpenAI-compat)
Hotkey	keyboard — ctrl+shift+d, ctrl+shift+a