Architecture

How Serpaix works.
From capture to recall.

Five layers — Capture, Process, Connect, Reason, Act — running on local SQLite (AES-256-GCM, OS keychain, 24-word BIP-39 recovery) with a 68-service self-host stack opening January 1, 2027. Cloud is opt-in, never required.

The 5-layer pipeline

Every note travels through 5 layers.

01

Capture

Voice memos (4 providers: Groq, Deepgram, AssemblyAI, local Whisper — fallback chain), ambient LTM daemons (active window, screenshot, clipboard every 5s), filesystem watcher (Rust notify), email-to-brain, YouTube, Obsidian/ENEX vault walk, API. All PII redacted in Rust before storage.

02

Process

OCR via Ollama vision → tesseract.js (offline). LLM extracts 11 entity types, generates embeddings, runs PII redaction, and creates summaries. All automatic, no manual tagging.

03

Connect

Auto-Linking Engine: entity overlap + semantic similarity (pgvector ≥0.65) + temporal proximity (±7 days). Graph clustering via Louvain. Confidence 0–1.0 on every edge. Desktop graph = vis-network; experimental WebGPU/3D on web.

04

Reason

Multi-hop RAG with graph traversal, contradiction detection, and temporal playback. Every answer cites its source memories. Voice Ask goes Whisper → local RAG → spoken answer, fully offline.

05

Act

Tasks auto-extracted from notes/meetings, linked to source. Pipes YAML+LLM automations (cron), MCP 22-tool memory layer for Claude/Cursor, calendars (Google/Outlook/ICS), Jira sync, publishing with custom domains.

Honest offline matrix

What works fully offline.

All core ops are local-first. Two items below are on the roadmap — we label them, not hide them.

FeatureOffline
Capture & transcription ✓
Filesystem watcher + OCR ✓
Graph + search (FTS5 + pgvector hybrid) ✓
AI chat (Ollama local) ✓
Voice Ask (Whisper → local RAG → TTS) ✓
Insights / contradictions / flashcards (FSRS-4.5) ✓
Cloud sync + E2E relay — On roadmap — flag false—
SIEM audit-log streaming — Roadmap—

Self-host stack (verified)

Root docker-compose.yml ships 68 services + 25 volumes (postgres/pgvector:pg16, qdrant v1.8.1, neo4j 5.20, nats 2.10, prometheus/grafana, temporal, redpanda — verified images). Lean launch stack is 6 services + 3 volumes (postgres, redis, backend, nlp-service, multimodal-service). Desktop itself uses none of Qdrant/Neo4j — graph is local SQLite + vis-network, vectors via pgvector/backend.

Source: audit/issues/FINDINGS_ADDENDUM_2026-08-23-site-crosscheck.md v2 (god-level verified).