📔 SNChat

A fully private, offline chatbot for your personal diary exported from Standard Notes. Ask natural-language questions about your journal — by topic, tag, or time period — and get answers grounded in your own entries. Nothing ever leaves your machine.

Built with Streamlit · LangChain · ChromaDB · Ollama.

Features

• 🔒 Fully private & offline — a local LLM (Ollama) and a local vector store (ChromaDB). No cloud, no API keys, no data leaving your computer.
• 🧠 Conversational memory — follow-up questions resolve against the chat history ("…in 2025?" → "and in 2026?") without losing the thread.
• 🏷️ Metadata-aware retrieval — filters by tag and date, not just semantic similarity. Multilingual tag matching lets you ask in English about non-English tags.
• 📅 Temporal alignment — chronological summaries and progressions ("how did my running develop?"), date ranges ("what did I do last week?", "how was my winter?" — including ranges across a year boundary), and "most recent N" queries ("summarize my last 10 runs").
• 🔎 Breadth that fits the question — narrow scopes return every matching entry; broad ones are summarized via map-reduce — instead of a fixed top-K cut-off.
• ⚡ Live & observable — answers stream token-by-token with a tokens/sec readout, and a sidebar gauge tracks context-window usage.

Setup

Requires Python 3.12+, uv, and a running Ollama.

# 1. Install dependencies
uv sync

# 2. Pull the local models (LLM + embeddings)
ollama pull gemma4:12b
ollama pull bge-m3

# 3. Run the app
uv run streamlit run app.py

Usage

1. Export your notes from Standard Notes as a .zip backup.
2. In the sidebar, upload the ZIP and click Process & Index Diary (one-time; the index persists locally in diary_vector_db/).
3. Ask away.

Your data

Each note's title must start with an ISO date (yyyy-mm-dd, optionally followed by a title) — all date filtering derives from this, and notes without a parseable date prefix are skipped. Tags come straight from the backup; multilingual tag synonyms are configurable in constants.py (TAG_ALIASES - keys must match the tags from the backup). Your diary and the built index are gitignored and never leave your machine.

Architecture

Three stages:

1. ingest a backup into a local vector store,
2. route each question into structured filters + a retrieval strategy,
3. generate a grounded answer.

All LLM and embedding calls go to a local Ollama; nothing leaves the machine.

sequenceDiagram
    actor User
    participant App as app.py
    participant Parser as parser.py
    participant Router as diary_query_router.py
    participant Gen as generation.py
    participant Chroma@{ "type" : "database" } as ChromaDB
    participant Ollama as Ollama

    rect rgba(68, 98, 144, 1)
    note over User, Ollama: 1. Ingest (once per backup)
    User->>App: Upload Standard Notes ZIP
    activate App
    App->>Parser: parse_standard_notes(zip)
    activate Parser
    Parser-->>App: Documents (notes + tags + date metadata)
    deactivate Parser
    App->>Ollama: Embed entries (bge-m3)
    activate Ollama
    Ollama-->>App: Vectors
    deactivate Ollama
    App->>Chroma: Persist to diary_vector_db/
    activate Chroma
    Chroma-->>App: Index ready
    deactivate Chroma
    deactivate App
    end

    rect rgba(77, 121, 77, 1)
    note over User, Ollama: 2. Route + retrieve (per question)
    User->>App: Ask a question
    activate App
    App->>Router: extract(question, history)
    activate Router
    Router->>Ollama: Parse into DiarySearchQuery (LLM)
    activate Ollama
    Ollama-->>Router: filters (date / tag / keywords / breadth)
    deactivate Ollama
    Router->>Chroma: retrieve() — similarity or metadata filter
    activate Chroma
    Chroma-->>Router: matching entries
    deactivate Chroma
    Router-->>App: retrieved entries
    deactivate Router
    end

    rect rgba(179, 113, 59, 1)
    note over User, Ollama: 3. Generate
    App->>Gen: plan_generation(entries, question, history)
    activate Gen
    Gen-->>App: messages (single-pass or map-reduce)
    deactivate Gen
    App->>Ollama: Stream answer (gemma4:12b)
    activate Ollama
    Ollama-->>App: tokens + usage metrics
    deactivate Ollama
    App-->>User: Streamed answer + citations
    deactivate App
    end

Loading

Docs

• CLAUDE.md — architecture & developer guide.
• docs/error_modes.md — known failure modes and their fixes.