OSpipe

RuVector-enhanced personal AI memory for Screenpipe

What is OSpipe?

Screenpipe is an open-source desktop application that continuously records your screen, audio, and UI interactions locally. It builds a searchable timeline of everything you see, hear, and do on your computer. Out of the box, Screenpipe stores its data in SQLite with FTS5 full-text indexing -- effective for keyword lookups, but limited to literal string matching. If you search for "auth discussion," you will not find a frame that says "we talked about login security."

OSpipe replaces Screenpipe's storage and search backend with the RuVector ecosystem -- a collection of 70+ Rust crates providing HNSW vector search, graph neural networks, attention mechanisms, delta-change tracking, and more. Instead of keyword matching, OSpipe embeds every captured frame into a high-dimensional vector space and performs approximate nearest neighbor search, delivering true semantic recall. A query like "what was that API we discussed in standup?" will surface the relevant audio transcription even if those exact words never appeared.

Everything stays local and private. OSpipe processes all data on-device with no cloud dependency. The safety gate automatically detects and redacts PII -- credit card numbers, Social Security numbers, and email addresses -- before content ever reaches the vector store. A cosine-similarity deduplication window prevents consecutive identical frames (like a static desktop) from bloating storage. Age-based quantization progressively compresses older embeddings from 32-bit floats down to 1-bit binary, cutting long-term memory usage by 97%.

OSpipe ships as a Rust crate, a TypeScript SDK, and a WASM library. It runs natively on Windows, macOS, and Linux, and can run entirely in the browser via WebAssembly at bundles as small as 11.8KB.

Ask your computer what you saw, heard, and did -- with semantic understanding.

Features

• Semantic Vector Search -- HNSW index via ruvector-core with 61us p50 query latency
• PII Safety Gate -- automatic redaction of credit card numbers, SSNs, and email addresses before storage
• Frame Deduplication -- cosine similarity sliding window eliminates near-duplicate captures
• Hybrid Search -- weighted combination of semantic vector similarity and keyword term overlap
• Query Router -- automatically routes queries to the optimal backend (Semantic, Keyword, Graph, Temporal, or Hybrid)
• WASM Support -- runs entirely in the browser with bundles from 11.8KB (micro) to 350KB (full)
• TypeScript SDK -- @ruvector/ospipe for Node.js and browser integration
• Configurable Quantization -- 4-tier age-based compression: f32 -> int8 -> product -> binary
• Cross-Platform -- native builds for Windows, macOS, Linux; WASM for browsers

Architecture

                         OSpipe Ingestion Pipeline
                         =========================

  Screenpipe -----> Capture -----> Safety Gate -----> Dedup -----> Embed -----> VectorStore
  (Screen/Audio/UI)  (CapturedFrame)  (PII Redaction)   (Cosine Window)  (HNSW)      |
                                                                                      |
                                                           Search Router <------------+
                                                           |    |    |    |    |
                                                        Semantic Keyword Graph Temporal Hybrid

Frames flow left to right through the ingestion pipeline. Each captured frame passes through:

1. Safety Gate -- PII detection and redaction; content may be allowed, redacted, or denied
2. Deduplication -- cosine similarity check against a sliding window of recent embeddings
3. Embedding -- text content is encoded into a normalized vector
4. Vector Store -- the embedding is indexed for approximate nearest neighbor retrieval

Queries enter through the Search Router, which analyzes the query string and dispatches to the optimal backend.

Quick Start

Rust

Add OSpipe to your Cargo.toml:

[dependencies]
ospipe = { path = "examples/OSpipe" }

Create a pipeline, ingest frames, and search:

use ospipe::config::OsPipeConfig;
use ospipe::pipeline::ingestion::IngestionPipeline;
use ospipe::capture::{CapturedFrame, CaptureSource, FrameContent, FrameMetadata};

fn main() -> ospipe::error::Result<()> {
    // Initialize with default configuration
    let config = OsPipeConfig::default();
    let mut pipeline = IngestionPipeline::new(config)?;

    // Ingest a screen capture
    let frame = CapturedFrame::new_screen(
        "Firefox",
        "Meeting Notes - Google Docs",
        "Discussion about authentication: we decided to use JWT with refresh tokens",
        0,
    );
    let result = pipeline.ingest(frame)?;
    println!("Ingest result: {:?}", result);

    // Ingest an audio transcription
    let audio = CapturedFrame::new_audio(
        "Built-in Microphone",
        "Let's revisit the login flow next sprint",
        Some("Alice"),
    );
    pipeline.ingest(audio)?;

    // Search semantically
    let query_embedding = pipeline.embedding_engine().embed("auth token discussion");
    let results = pipeline.vector_store().search(&query_embedding, 5)?;

    for hit in &results {
        println!("Score: {:.4} | {:?}", hit.score, hit.metadata);
    }

    // Print pipeline statistics
    let stats = pipeline.stats();
    println!(
        "Ingested: {} | Deduped: {} | Denied: {} | Redacted: {}",
        stats.total_ingested, stats.total_deduplicated,
        stats.total_denied, stats.total_redacted
    );

    Ok(())
}

TypeScript

import { OsPipe } from "@ruvector/ospipe";

const client = new OsPipe({ baseUrl: "http://localhost:3030" });

// Ingest a captured frame
await client.ingest({
  source: "screen",
  app: "Chrome",
  window: "Jira Board",
  content: "Sprint 14 planning: migrate auth to OAuth2",
});

// Semantic search
const results = await client.queryRuVector(
  "what did I discuss in the meeting about authentication?"
);

for (const hit of results) {
  console.log(`[${hit.score.toFixed(3)}] ${hit.metadata.text}`);
}

WASM (Browser)

import { OsPipeWasm } from "@ruvector/ospipe-wasm";

// Initialize with 384-dimensional embeddings
const pipe = new OsPipeWasm(384);

// Embed and insert content
const embedding = pipe.embed_text("meeting notes about auth migration to OAuth2");
pipe.insert("frame-001", embedding, '{"app":"Chrome","window":"Jira"}', Date.now());

// Embed a query and search
const queryEmbedding = pipe.embed_text("what was the auth discussion about?");
const results = pipe.search(queryEmbedding, 5);
console.log("Results:", results);

// Safety check before storage
const safety = pipe.safety_check("my card is 4111-1111-1111-1111");
console.log("Safety:", safety); // "deny"

// Query routing
const route = pipe.route_query("what happened yesterday?");
console.log("Route:", route); // "Temporal"

// Pipeline statistics
console.log("Stats:", pipe.stats());

Comparison: Screenpipe vs OSpipe

RuVector Crate Integration

Configuration

use ospipe::config::OsPipeConfig;

let config = OsPipeConfig::default();
// config.data_dir        = "~/.ospipe"
// config.capture         = CaptureConfig { ... }
// config.storage         = StorageConfig { ... }
// config.search          = SearchConfig { ... }
// config.safety          = SafetyConfig { ... }

use ospipe::config::*;
use std::path::PathBuf;

let config = OsPipeConfig {
    data_dir: PathBuf::from("/var/lib/ospipe"),
    capture: CaptureConfig {
        fps: 0.5,
        audio_chunk_secs: 60,
        excluded_apps: vec![
            "1Password".into(),
            "Signal".into(),
            "Bitwarden".into(),
        ],
        skip_private_windows: true,
    },
    storage: StorageConfig {
        embedding_dim: 768,      // Use a larger model
        hnsw_m: 48,              // More connections for better recall
        hnsw_ef_construction: 400,
        hnsw_ef_search: 200,
        dedup_threshold: 0.98,   // Stricter deduplication
        ..Default::default()
    },
    search: SearchConfig {
        default_k: 20,
        hybrid_weight: 0.8,      // Lean more toward semantic
        mmr_lambda: 0.6,
        rerank_enabled: true,
    },
    safety: SafetyConfig {
        pii_detection: true,
        credit_card_redaction: true,
        ssn_redaction: true,
        custom_patterns: vec![
            "INTERNAL_ONLY".into(),
            "CONFIDENTIAL".into(),
        ],
    },
};

Safety Gate

4111111111111111       -> [CC_REDACTED]
4111 1111 1111 1111    -> [CC_REDACTED]
4111-1111-1111-1111    -> [CC_REDACTED]

123-45-6789            -> [SSN_REDACTED]

user@example.com       -> [EMAIL_REDACTED]
admin@company.org      -> [EMAIL_REDACTED]

let config = SafetyConfig {
    custom_patterns: vec!["TOP_SECRET".to_string(), "CLASSIFIED".to_string()],
    ..Default::default()
};

pipe.safety_check("my card is 4111-1111-1111-1111"); // "deny"
pipe.safety_check("set password to foo123");          // "redact"
pipe.safety_check("the weather is nice today");       // "allow"

Advanced Configuration

// worker.js
import { OsPipeWasm } from "@ruvector/ospipe-wasm";

const pipe = new OsPipeWasm(384);

self.onmessage = (event) => {
  const { type, payload } = event.data;

  switch (type) {
    case "insert":
      const emb = pipe.embed_text(payload.text);
      pipe.insert(payload.id, emb, JSON.stringify(payload.metadata), Date.now());
      self.postMessage({ type: "inserted", id: payload.id });
      break;

    case "search":
      const queryEmb = pipe.embed_text(payload.query);
      const results = pipe.search(queryEmb, payload.k || 10);
      self.postMessage({ type: "results", data: results });
      break;
  }
};

Cross-Origin-Opener-Policy: same-origin
Cross-Origin-Embedder-Policy: require-corp

# Native (current platform)
cargo build -p ospipe --release

# WASM (browser)
cargo build -p ospipe --target wasm32-unknown-unknown --release

# Generate JS bindings
wasm-pack build examples/OSpipe --target web --release

# Windows (cross-compile)
cross build -p ospipe --target x86_64-pc-windows-gnu --release

# macOS ARM (cross-compile)
cross build -p ospipe --target aarch64-apple-darwin --release

# macOS Intel (cross-compile)
cross build -p ospipe --target x86_64-apple-darwin --release

# Linux ARM (cross-compile)
cross build -p ospipe --target aarch64-unknown-linux-gnu --release

use ospipe::config::{StorageConfig, QuantizationTier, QuantizationMethod};

let storage = StorageConfig {
    quantization_tiers: vec![
        QuantizationTier { age_hours: 0,    method: QuantizationMethod::None },
        QuantizationTier { age_hours: 12,   method: QuantizationMethod::Scalar },
        QuantizationTier { age_hours: 72,   method: QuantizationMethod::Product },
        QuantizationTier { age_hours: 360,  method: QuantizationMethod::Binary },
    ],
    ..Default::default()
};

API Reference

Rust API

Core Types

Pipeline

Storage

Search

Safety

WASM API (OsPipeWasm)

Testing

# Run all 56 tests
cargo test -p ospipe

# Run with verbose output
cargo test -p ospipe -- --nocapture

# Run only integration tests
cargo test -p ospipe --test integration

# Run only unit tests (embedding, WASM helpers)
cargo test -p ospipe --lib

# Build for WASM (verify compilation)
cargo build -p ospipe --target wasm32-unknown-unknown

# Build with wasm-pack for JS bindings
wasm-pack build examples/OSpipe --target web

Test Coverage

Related

• ADR: OSpipe Screenpipe Integration -- Architecture Decision Record with full design rationale
• Screenpipe -- Open-source local-first desktop recording + AI memory
• RuVector -- 70+ Rust crates for vector search, graph neural networks, and attention mechanisms
• @ruvector/ospipe -- TypeScript SDK (npm)
• @ruvector/ospipe-wasm -- WASM package (npm)

License

Licensed under either of:

• MIT License (LICENSE-MIT or http://opensource.org/licenses/MIT)
• Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)

at your option.