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:

Safety Gate -- PII detection and redaction; content may be allowed, redacted, or denied
Deduplication -- cosine similarity check against a sliding window of recent embeddings
Embedding -- text content is encoded into a normalized vector
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

Feature	Screenpipe (FTS5)	OSpipe (RuVector)
Search Type	Keyword (FTS5)	Semantic + Keyword + Graph + Temporal
Search Latency	~1ms (FTS5)	61us (HNSW p50)
Content Relations	None	Knowledge Graph (Cypher)
Temporal Analysis	Basic SQL	Delta-behavior tracking
PII Protection	Basic	Credit card, SSN, email redaction
Deduplication	None	Cosine similarity sliding window
Browser Support	None	WASM (11.8KB - 350KB)
Quantization	None	4-tier age-based (f32 -> binary)
Privacy	Local-first	Local-first + PII redaction
Query Routing	None	Auto-routes to optimal backend
Hybrid Search	None	Weighted semantic + keyword fusion
Metadata Filtering	SQL WHERE	App, time range, content type, monitor

RuVector Crate Integration

RuVector Crate	OSpipe Usage	Status
`ruvector-core`	HNSW vector storage and nearest neighbor search	Integrated
`ruvector-filter`	Metadata filtering (app, time, content type)	Integrated
`ruvector-cluster`	Frame deduplication via cosine similarity	Integrated
`ruvector-delta-core`	Change tracking and delta-behavior analysis	Integrated
`ruvector-router-core`	Query routing to optimal search backend	Integrated
`cognitum-gate-kernel`	AI safety gate decisions (allow/redact/deny)	Integrated
`ruvector-graph`	Knowledge graph for entity relationships	Phase 2
`ruvector-attention`	Content prioritization and relevance weighting	Phase 3
`ruvector-gnn`	Learned search improvement via graph neural nets	Phase 3
`ruqu-algorithms`	Quantum-inspired search acceleration	Phase 4

Configuration

Full Configuration Reference

`OsPipeConfig`

Top-level configuration with nested subsystem configs. All fields have sensible defaults.

use ospipe::config::OsPipeConfig;

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

`CaptureConfig`

Field	Type	Default	Description
`fps`	`f32`	`1.0`	Frames per second for screen capture
`audio_chunk_secs`	`u32`	`30`	Duration of audio chunks in seconds
`excluded_apps`	`Vec<String>`	`["1Password", "Keychain Access"]`	Applications excluded from capture
`skip_private_windows`	`bool`	`true`	Skip windows marked as private/incognito

`StorageConfig`

Field	Type	Default	Description
`embedding_dim`	`usize`	`384`	Dimensionality of embedding vectors
`hnsw_m`	`usize`	`32`	HNSW M parameter (max connections per layer)
`hnsw_ef_construction`	`usize`	`200`	HNSW ef_construction (index build quality)
`hnsw_ef_search`	`usize`	`100`	HNSW ef_search (query-time accuracy)
`dedup_threshold`	`f32`	`0.95`	Cosine similarity threshold for deduplication
`quantization_tiers`	`Vec<QuantizationTier>`	4 tiers (see below)	Age-based quantization schedule

`SearchConfig`

Field	Type	Default	Description
`default_k`	`usize`	`10`	Default number of results to return
`hybrid_weight`	`f32`	`0.7`	Semantic vs keyword weight (1.0 = pure semantic, 0.0 = pure keyword)
`mmr_lambda`	`f32`	`0.5`	MMR diversity vs relevance tradeoff
`rerank_enabled`	`bool`	`false`	Whether to enable result reranking

`SafetyConfig`

Field	Type	Default	Description
`pii_detection`	`bool`	`true`	Enable PII detection (emails)
`credit_card_redaction`	`bool`	`true`	Enable credit card number redaction
`ssn_redaction`	`bool`	`true`	Enable SSN redaction
`custom_patterns`	`Vec<String>`	`[]`	Custom substring patterns that trigger denial

Example: Custom Configuration

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

PII Detection Details

The safety gate inspects all captured content before it enters the ingestion pipeline. It operates in three modes:

Safety Decisions

Decision	Behavior	When
`Allow`	Content stored as-is	No sensitive patterns detected
`AllowRedacted(String)`	Content stored with PII replaced by tokens	PII detected, redaction enabled
`Deny { reason }`	Content rejected, not stored	Custom deny pattern matched

Detected PII Patterns

Credit Cards -- sequences of 13-16 digits (with optional spaces or dashes):

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

Social Security Numbers -- XXX-XX-XXXX format:

123-45-6789            -> [SSN_REDACTED]

Email Addresses -- word@domain.tld patterns:

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

Custom Patterns -- configurable substring deny list. When a custom pattern is matched, the entire frame is denied (not just redacted):

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

WASM Safety API

The WASM bindings expose a simplified safety classifier:

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"

The WASM classifier also detects sensitive keywords: password, secret, api_key, api-key, apikey, token, private_key, private-key.

Advanced Configuration

WASM Deployment

Bundle Tiers

OSpipe provides four WASM bundle sizes depending on which features you need:

Tier	Size	Features
Micro	11.8KB	Embedding + vector search only
Standard	225KB	Full pipeline (embed, insert, search, filtered search)
Full	350KB	+ deduplication + safety gate + query routing
AI	2.5MB	+ on-device neural inference (ONNX)

Web Worker Setup

For best performance, run OSpipe in a Web Worker to avoid blocking the main thread:

// 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;
  }
};

SharedArrayBuffer

For multi-threaded WASM (e.g., parallel batch embedding), set the required headers:

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

Cross-Platform Build

Build Targets

# 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

Conditional Compilation

OSpipe uses conditional compilation to separate native and WASM dependencies:

Native (cfg(not(target_arch = "wasm32"))) -- links against ruvector-core, ruvector-filter, ruvector-cluster, ruvector-delta-core, ruvector-router-core, and cognitum-gate-kernel
WASM (cfg(target_arch = "wasm32")) -- uses wasm-bindgen, js-sys, serde-wasm-bindgen, and getrandom with the js feature

The src/wasm/helpers.rs module contains pure Rust functions (cosine similarity, hash embedding, safety classification, query routing) that compile on all targets and are tested natively.

Quantization Tiers

OSpipe progressively compresses older embeddings to reduce long-term storage costs. The default quantization schedule:

Age	Method	Bits/Dim	Memory vs f32	Description
0 hours	None (f32)	32	100%	Full precision for recent content
24 hours	Scalar (int8)	8	25%	Minimal quality loss, 4x compression
1 week	Product	~2	~6%	Codebook-based compression
30 days	Binary	1	3%	Single bit per dimension, 97% savings

Custom Tiers

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()
};

Memory Estimate

For 1 million frames at 384 dimensions:

Tier	Bytes/Vector	Total (1M vectors)
f32	1,536	1.43 GB
int8	384	366 MB
Product	~96	~91 MB
Binary	48	46 MB

With the default age distribution (most content aging past 30 days), long-term average storage is approximately 50-80 MB per million frames.

API Reference

Rust API

Core Types

Type	Module	Description
`OsPipeConfig`	`config`	Top-level configuration
`CaptureConfig`	`config`	Capture subsystem settings
`StorageConfig`	`config`	HNSW and quantization settings
`SearchConfig`	`config`	Search weights and defaults
`SafetyConfig`	`config`	PII detection toggles
`CapturedFrame`	`capture`	A captured screen/audio/UI frame
`CaptureSource`	`capture`	Source enum: `Screen`, `Audio`, `Ui`
`FrameContent`	`capture`	Content enum: `OcrText`, `Transcription`, `UiEvent`
`FrameMetadata`	`capture`	Metadata (app, window, monitor, confidence, language)
`OsPipeError`	`error`	Unified error type

Pipeline

Type / Function	Module	Description
`IngestionPipeline::new(config)`	`pipeline::ingestion`	Create a new pipeline
`IngestionPipeline::ingest(frame)`	`pipeline::ingestion`	Ingest a single frame
`IngestionPipeline::ingest_batch(frames)`	`pipeline::ingestion`	Ingest multiple frames
`IngestionPipeline::stats()`	`pipeline::ingestion`	Get ingestion statistics
`IngestResult`	`pipeline::ingestion`	Enum: `Stored`, `Deduplicated`, `Denied`
`PipelineStats`	`pipeline::ingestion`	Counters for ingested/deduped/denied/redacted
`FrameDeduplicator`	`pipeline::dedup`	Cosine similarity sliding window

Storage

Type / Function	Module	Description
`VectorStore::new(config)`	`storage::vector_store`	Create a new vector store
`VectorStore::insert(frame, embedding)`	`storage::vector_store`	Insert a frame with its embedding
`VectorStore::search(query, k)`	`storage::vector_store`	Top-k nearest neighbor search
`VectorStore::search_filtered(query, k, filter)`	`storage::vector_store`	Search with metadata filters
`SearchResult`	`storage::vector_store`	Result with id, score, metadata
`SearchFilter`	`storage::vector_store`	Filter by app, time range, content type, monitor
`StoredEmbedding`	`storage::vector_store`	Stored vector with metadata and timestamp
`EmbeddingEngine::new(dim)`	`storage::embedding`	Create an embedding engine
`EmbeddingEngine::embed(text)`	`storage::embedding`	Generate a normalized embedding
`EmbeddingEngine::batch_embed(texts)`	`storage::embedding`	Batch embedding generation
`cosine_similarity(a, b)`	`storage::embedding`	Cosine similarity between two vectors

Search

Type / Function	Module	Description
`QueryRouter::new()`	`search::router`	Create a query router
`QueryRouter::route(query)`	`search::router`	Route a query to optimal backend
`QueryRoute`	`search::router`	Enum: `Semantic`, `Keyword`, `Graph`, `Temporal`, `Hybrid`
`HybridSearch::new(weight)`	`search::hybrid`	Create a hybrid search with semantic weight
`HybridSearch::search(store, query, emb, k)`	`search::hybrid`	Combined semantic + keyword search

Safety

Type / Function	Module	Description
`SafetyGate::new(config)`	`safety`	Create a safety gate
`SafetyGate::check(content)`	`safety`	Check content, return safety decision
`SafetyGate::redact(content)`	`safety`	Redact and return cleaned content
`SafetyDecision`	`safety`	Enum: `Allow`, `AllowRedacted(String)`, `Deny { reason }`

WASM API (`OsPipeWasm`)

Method	Parameters	Returns	Description
`new(dimension)`	`usize`	`OsPipeWasm`	Constructor
`insert(id, embedding, metadata, timestamp)`	`&str, &[f32], &str, f64`	`Result<(), JsValue>`	Insert a frame
`search(query_embedding, k)`	`&[f32], usize`	`JsValue` (JSON array)	Semantic search
`search_filtered(query_embedding, k, start, end)`	`&[f32], usize, f64, f64`	`JsValue` (JSON array)	Time-filtered search
`is_duplicate(embedding, threshold)`	`&[f32], f32`	`bool`	Deduplication check
`embed_text(text)`	`&str`	`Vec<f32>`	Hash-based text embedding
`batch_embed(texts)`	`JsValue` (Array)	`JsValue` (Array)	Batch text embedding
`safety_check(content)`	`&str`	`String`	Returns "allow", "redact", or "deny"
`route_query(query)`	`&str`	`String`	Returns "Semantic", "Keyword", "Graph", or "Temporal"
`len()`	--	`usize`	Number of stored embeddings
`stats()`	--	`String` (JSON)	Pipeline statistics

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

Test Category	Count	Module
Configuration	2	`tests/integration.rs`
Capture frames	3	`tests/integration.rs`
Embedding engine	6	`src/storage/embedding.rs`
Vector store	4	`tests/integration.rs`
Deduplication	2	`tests/integration.rs`
Safety gate	6	`tests/integration.rs`
Query routing	4	`tests/integration.rs`
Hybrid search	2	`tests/integration.rs`
Ingestion pipeline	5	`tests/integration.rs`
Cosine similarity	3	`tests/integration.rs`
WASM helpers	18	`src/wasm/helpers.rs`
Total	56

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.

FilesExpand file tree

README.md

Latest commit

History