Feat/ruvector postgres v2#82
Merged
Merged
Conversation
Complete specification for RuVector Postgres v2 with: Architecture: - PostgreSQL extension (pgrx) with hybrid architecture - SQL handles ACID/joins, RuVector engine handles vectors/graphs/learning - Backward compatible with pgvector SQL surface - Shared memory IPC with bounded contracts (64KB inline, 16MB shared) 4-Phase Implementation: - Phase 1: pgvector-compatible search (1a: function-based, 1b: Index AM) - Phase 2: Tiered storage with compression and exactness GUC - Phase 3: Graph engine with Cypher and SQL join keys - Phase 4: Dynamic mincut integrity gating (key differentiator) Key Technical Details: - lambda_cut: Minimum cut value via Stoer-Wagner (PRIMARY integrity metric) - lambda2: Algebraic connectivity (OPTIONAL drift signal) - DIFFERENT from mincut! - Contracted operational graph (~1000 nodes) - never compute on full similarity graph - Hysteresis model with consecutive samples and cooldown - Operation risk classification (Low/Medium/High) - MVCC visibility with incremental paging API - WAL replay with idempotency and LSN ordering - Partition map versioning and epoch fencing for cluster mode Files: - 00-overview.md: Architecture, consistency contract, benchmark spec - 01-sql-schema.md: SQL schema and types - 02-background-workers.md: IPC contract, mincut worker - 03-index-access-methods.md: Index AM specification - 04-integrity-events.md: Events, hysteresis, operation classes - 05-phase1-pgvector-compat.md: Phase 1a/1b incremental path - 06-phase2-tiered-storage.md: Tiered storage with GUC exactness - 07-phase3-graph-cypher.md: Graph engine with SQL joins - 08-phase4-integrity-control.md: Mincut gating with Stoer-Wagner - 09-migration-guide.md: Migration from pgvector - 10-consistency-replication.md: Consistency and replication model 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace technical executive summary with clear explanation of why RuVector matters: - From symptom monitoring to causal monitoring - Mincut as leading indicator, not metric - Algorithm becomes control signal (control plane, not analytics) - Failure mode class change: cascading → graceful degradation - Explainable operations via witness edges Key message: "We're not making vector search faster. We're making vector infrastructure survivable." 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Three high-impact additions to RuVector Postgres v2: ## 11-hybrid-search.md - BM25 + Vector Fusion - Single query combines semantic and keyword search - Proper BM25 implementation (not just ts_rank) - Fusion algorithms: RRF (default), linear, learned - Integrity-aware degradation (stress → single branch) - Parallel branch execution - GUC configuration ## 12-multi-tenancy.md - First-Class Tenant Isolation - SET ruvector.tenant_id for transparent scoping - Isolation levels: shared, partition, dedicated - Automatic promotion based on vector count - Per-tenant integrity (stress in one doesn't affect others) - Per-tenant contracted graphs - Resource quotas and rate limiting - Fair scheduling (no noisy neighbors) - RLS integration for defense in depth ## 13-self-healing.md - Automated Remediation - Completes the control loop: sensor → actuator - Problem classification from witness edges: - Hotspot congestion - Centroid skew - Replication lag - Maintenance contention - Index fragmentation - Memory pressure - Built-in strategies: - Rebalance partitions - Pause maintenance jobs - Throttle ingestion - Scale read replicas (K8s) - Compact fragmented indexes - Safety: reversible actions, blast radius limits - Learning: outcome tracking, strategy weight updates - The key insight: "We built the sensor. Now we build the actuator." 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
…ures Comprehensive intelligence system for Claude Code hooks: Core Features (v2): - VectorMemory with @ruvector/core native HNSW (150x faster) - Hyperbolic distance (Poincaré ball) for hierarchical embeddings - ReasoningBank with Q-learning and pattern decay (7-day half-life) - Confidence Calibration tracking (predicted vs actual accuracy) - A/B Testing with 10% holdout for measuring intelligence lift - Feedback Loop for tracking suggestion follow-through - Active Learning for identifying uncertain states v3 Improvements: - Error Pattern Learning (Rust E0xxx, TypeScript TSxxxx, npm errors) - File Sequence Learning (tracks which files are edited together) - Test Suggestion Triggers (suggests cargo test after source edits) - Hive-Mind swarm coordination (11 agents, 38 edges) Pretrained from memory.db: - 7,697 commands processed - 4,023 vector memories - 117 Q-table states with decay metadata - 8,520 calibration samples Anti-overfitting measures: - Q-values capped at 0.8, floored at -0.5 - Decaying learning rate: 0.3/sqrt(count) - Pattern decay with timestamps 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Three critical bugs were preventing the intelligence layer from using
learned patterns:
1. State format mismatch: CLI used spaces ("editing rs in project")
but Q-table used underscores ("edit_rs_in_project")
- Fixed in cli.js: all states now use underscore format
2. stateKey() hyphen normalization: Function converted hyphens to
underscores, but Q-table keys had hyphens (e.g. "ruvector-core")
- Fixed regex: /[^a-z0-9-]+/g preserves hyphens
3. A/B testing control group: 10% random sessions ignored learning
- Reduced holdout to 5% with persistent session assignment
- Added INTELLIGENCE_MODE=treatment env override for development
Result: Agent recommendations now show 80% confidence for Rust files
using learned Q-values, instead of 0% with random selection.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Critical fix: PreToolUse hooks were running in background (&) which meant Claude never saw the intelligence output. Now: - PreToolUse: Foreground execution (Claude sees guidance) - pre-edit: Shows recommended agent + confidence + similar edits - pre-command: Shows command patterns + suggestions - Added 3s timeout to prevent blocking - PostToolUse: Background execution (async learning) - post-edit: Records success/failure, learns patterns - post-command: Captures errors, updates Q-values - SessionStart: New hook shows learned patterns at session start - Displays pattern count, memory stats - Shows top 3 learned state-action pairs with Q-values Claude now receives self-learning guidance like: "🧠 Intelligence Analysis: 📁 ruvector-core/lib.rs 🤖 Recommended: rust-developer (80% confidence) 📚 3 similar past edits found" 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
ruvnet
added a commit
that referenced
this pull request
Feb 20, 2026
* feat(postgres): Add RuVector Postgres v2 implementation plan Complete specification for RuVector Postgres v2 with: Architecture: - PostgreSQL extension (pgrx) with hybrid architecture - SQL handles ACID/joins, RuVector engine handles vectors/graphs/learning - Backward compatible with pgvector SQL surface - Shared memory IPC with bounded contracts (64KB inline, 16MB shared) 4-Phase Implementation: - Phase 1: pgvector-compatible search (1a: function-based, 1b: Index AM) - Phase 2: Tiered storage with compression and exactness GUC - Phase 3: Graph engine with Cypher and SQL join keys - Phase 4: Dynamic mincut integrity gating (key differentiator) Key Technical Details: - lambda_cut: Minimum cut value via Stoer-Wagner (PRIMARY integrity metric) - lambda2: Algebraic connectivity (OPTIONAL drift signal) - DIFFERENT from mincut! - Contracted operational graph (~1000 nodes) - never compute on full similarity graph - Hysteresis model with consecutive samples and cooldown - Operation risk classification (Low/Medium/High) - MVCC visibility with incremental paging API - WAL replay with idempotency and LSN ordering - Partition map versioning and epoch fencing for cluster mode Files: - 00-overview.md: Architecture, consistency contract, benchmark spec - 01-sql-schema.md: SQL schema and types - 02-background-workers.md: IPC contract, mincut worker - 03-index-access-methods.md: Index AM specification - 04-integrity-events.md: Events, hysteresis, operation classes - 05-phase1-pgvector-compat.md: Phase 1a/1b incremental path - 06-phase2-tiered-storage.md: Tiered storage with GUC exactness - 07-phase3-graph-cypher.md: Graph engine with SQL joins - 08-phase4-integrity-control.md: Mincut gating with Stoer-Wagner - 09-migration-guide.md: Migration from pgvector - 10-consistency-replication.md: Consistency and replication model 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * docs(postgres): Rewrite v2 overview with compelling framing Replace technical executive summary with clear explanation of why RuVector matters: - From symptom monitoring to causal monitoring - Mincut as leading indicator, not metric - Algorithm becomes control signal (control plane, not analytics) - Failure mode class change: cascading → graceful degradation - Explainable operations via witness edges Key message: "We're not making vector search faster. We're making vector infrastructure survivable." 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * feat(postgres): Add hybrid search, multi-tenancy, and self-healing specs Three high-impact additions to RuVector Postgres v2: ## 11-hybrid-search.md - BM25 + Vector Fusion - Single query combines semantic and keyword search - Proper BM25 implementation (not just ts_rank) - Fusion algorithms: RRF (default), linear, learned - Integrity-aware degradation (stress → single branch) - Parallel branch execution - GUC configuration ## 12-multi-tenancy.md - First-Class Tenant Isolation - SET ruvector.tenant_id for transparent scoping - Isolation levels: shared, partition, dedicated - Automatic promotion based on vector count - Per-tenant integrity (stress in one doesn't affect others) - Per-tenant contracted graphs - Resource quotas and rate limiting - Fair scheduling (no noisy neighbors) - RLS integration for defense in depth ## 13-self-healing.md - Automated Remediation - Completes the control loop: sensor → actuator - Problem classification from witness edges: - Hotspot congestion - Centroid skew - Replication lag - Maintenance contention - Index fragmentation - Memory pressure - Built-in strategies: - Rebalance partitions - Pause maintenance jobs - Throttle ingestion - Scale read replicas (K8s) - Compact fragmented indexes - Safety: reversible actions, blast radius limits - Learning: outcome tracking, strategy weight updates - The key insight: "We built the sensor. Now we build the actuator." 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * feat(intelligence): Add self-learning intelligence layer with v3 features Comprehensive intelligence system for Claude Code hooks: Core Features (v2): - VectorMemory with @ruvector/core native HNSW (150x faster) - Hyperbolic distance (Poincaré ball) for hierarchical embeddings - ReasoningBank with Q-learning and pattern decay (7-day half-life) - Confidence Calibration tracking (predicted vs actual accuracy) - A/B Testing with 10% holdout for measuring intelligence lift - Feedback Loop for tracking suggestion follow-through - Active Learning for identifying uncertain states v3 Improvements: - Error Pattern Learning (Rust E0xxx, TypeScript TSxxxx, npm errors) - File Sequence Learning (tracks which files are edited together) - Test Suggestion Triggers (suggests cargo test after source edits) - Hive-Mind swarm coordination (11 agents, 38 edges) Pretrained from memory.db: - 7,697 commands processed - 4,023 vector memories - 117 Q-table states with decay metadata - 8,520 calibration samples Anti-overfitting measures: - Q-values capped at 0.8, floored at -0.5 - Decaying learning rate: 0.3/sqrt(count) - Pattern decay with timestamps 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * fix(intelligence): Fix Q-table lookups - learning now has real effect Three critical bugs were preventing the intelligence layer from using learned patterns: 1. State format mismatch: CLI used spaces ("editing rs in project") but Q-table used underscores ("edit_rs_in_project") - Fixed in cli.js: all states now use underscore format 2. stateKey() hyphen normalization: Function converted hyphens to underscores, but Q-table keys had hyphens (e.g. "ruvector-core") - Fixed regex: /[^a-z0-9-]+/g preserves hyphens 3. A/B testing control group: 10% random sessions ignored learning - Reduced holdout to 5% with persistent session assignment - Added INTELLIGENCE_MODE=treatment env override for development Result: Agent recommendations now show 80% confidence for Rust files using learned Q-values, instead of 0% with random selection. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * fix(hooks): Display intelligence guidance to Claude in foreground Critical fix: PreToolUse hooks were running in background (&) which meant Claude never saw the intelligence output. Now: - PreToolUse: Foreground execution (Claude sees guidance) - pre-edit: Shows recommended agent + confidence + similar edits - pre-command: Shows command patterns + suggestions - Added 3s timeout to prevent blocking - PostToolUse: Background execution (async learning) - post-edit: Records success/failure, learns patterns - post-command: Captures errors, updates Q-values - SessionStart: New hook shows learned patterns at session start - Displays pattern count, memory stats - Shows top 3 learned state-action pairs with Q-values Claude now receives self-learning guidance like: "🧠 Intelligence Analysis: 📁 ruvector-core/lib.rs 🤖 Recommended: rust-developer (80% confidence) 📚 3 similar past edits found" 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
drdave-flexnetos
referenced
this pull request
in FlexNetOS/meta-ruvector
Jun 26, 2026
Wired dep-fix clusters: - Add crates/ruvector-hyperbolic-hnsw + wasm to members (nalgebra 0.34.1, approx, WASM extras) - Add examples/edge-net to members (crypto deps: ed25519-dalek, x25519-dalek, aes-gcm, sha2, argon2, zeroize, hex, base64, rustc-hash, typed-arena, string_cache) - Added ~50 workspace dependencies covering: DDS/ROS2 middleware (zenoh, rustdds, cdr), vibecast/sevensense (~30 deps), edge-net crypto, WASM extras Added workspace.package fields (homepage, documentation, description, keywords, categories) for agentic-robotics readiness. Updated tokio to "full" features (needed by robotics/vibecast clusters). Added bincode1 alias for 1.x lineage compatibility. Build: 0 errors, 564 crates (was 548), workspace compiles green
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This pull request introduces a new suite of developer productivity and intelligence hooks for the RuVector project, focused on Rust and WASM crate workflows. The additions automate and standardize pre- and post-edit checks, benchmarking, WASM size validation, and context extraction for intelligent code assistance. Additionally, a detailed plan for further intelligence system improvements is documented.
The most important changes are:
Developer Workflow Automation:
.claude/hooks/rust-check.shto provide a Rust-specific pre-edit hook that identifies the crate, summarizes its purpose, and runscargo checkwith context-aware output for the developer..claude/hooks/post-rust-edit.shas a post-edit hook for Rust files, running formatting checks, clippy, optional benchmarks, and recording edit metrics for later analysis.Benchmarking and WASM Validation:
.claude/hooks/bench-runner.shto automate running and storing results for criterion benchmarks across core, mincut, attention, and graph crates, supporting both targeted and all-crate runs..claude/hooks/wasm-size-check.shto check WASM bundle sizes for all relevant crates, ensuring compliance with the <3MB gzipped target and reporting results clearly.Intelligent Context Extraction:
.claude/hooks/crate-context.shto output JSON-formatted context for any file, including related tests, examples, documentation, key dependencies, and suggested commands, supporting smarter code assistance.Planning and Documentation:
.claude/intelligence/IMPROVEMENTS.mdoutlining a prioritized roadmap for enhancing the intelligence system, including error pattern learning, file sequence tracking, smarter test/build suggestions, and more.