meb

package module
v0.7.0 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: Aug 27, 2026 License: Apache-2.0 Imports: 27 Imported by: 0

README

meb

CI

Mangle Extension for Badger — an embedded knowledge graph database combining triple-store semantics with Hybrid (FWHT + Block-wise) vector search.

Features

  • Triple Store: Subject-Predicate-Object with dual SPO/OPS indexing (25-byte keys)
  • Multi-Topic Isolation: 24-bit TopicID bit-packing enables 16M namespaces without a Graph column
  • ANN Indexes: IVF-PQ (coarse K-means + residual PQ) and HNSW (hierarchical navigable small world) for O(log N) vector search
  • Badger-Native Vector Store: Compressed vectors stored in BadgerDB with mmap cache layer — disk-scaled, O(k) memory per search
  • SIMD ADC Accumulation: Asymmetric distance computation with scalar, AVX2 (amd64), and NEON (arm64) kernels; runtime dispatch via golang.org/x/sys/cpu
  • Dual-Path Search: Hot queries hit parallel mmap (~500K vectors/sec); cold start streams from Badger
  • LSM-Level Topic Filtering: SearchInTopic() uses Badger prefix scan — zero I/O on unrelated topics
  • Hybrid Vector Compression: FWHT preconditioning + block-wise 4/8-bit quantization preserving full 1536 dimensions
  • Cauchy-Schwarz Pruning: Per-block L2 norm suffix sums for early termination during vector search
  • Zero-Copy Streaming: Go 1.23+ iter.Seq2 for constant-memory scan operations
  • Unified BadgerDB: Single database for graph, dictionary, vectors, and content — enables cross-subsystem transactions
  • Cross-Subsystem Transactions: Opt-in View()/Update() API for atomic multi-operation writes (graph + vector + content + dictionary)
  • Datalog Integration: Mangle factstore.FactStore interface for symbolic reasoning
  • Neuro-Symbolic Search: Hybrid vector + LFTJ graph query builder with streaming joins
  • Rule-Based Optimizer (RBO): Auto-selects GraphFirst, VectorFirst, IntersectionFirst, PureVectorSearch, or PureLFTJ strategy per query
  • Cost-Based Optimizer (CBO): Greedy smallest-first join ordering with plan cache
  • Streaming Execution: Concurrent dense/graph path merge with early termination via context cancellation
  • Circuit Breaker: Configurable query timeout protection with push telemetry
  • Hybrid WAL Approach: Transactions use BadgerDB SyncWrites: true for durability; WriteBatch path uses WAL v2 (CRC32C) for crash recovery
  • Deterministic LFTJ Joins: Canonical ordering + ExecuteOrdered for reproducible multi-way join results
  • Sharded Dictionary: Configurable lock-striping for high-concurrent ingestion
  • Soft Delete: HNSW tombstone keys + Compact() graph rebuild
  • S2 Compression: Fast content storage with Snappy-compatible compression

Quick Start

package main

import (
    "context"
    "log"
    "github.com/duynguyendang/meb"
    "github.com/duynguyendang/meb/store"
)

func main() {
    ctx := context.Background()
    cfg := store.DefaultConfig("./meb-data")
    s, err := meb.NewMEBStore(cfg)
    if err != nil {
        log.Fatal(err)
    }
    defer s.Close()

    // Add facts
    s.AddFact(meb.NewFact("Alice", "knows", "Bob"))
    s.AddFact(meb.NewFact("Alice", "works_at", "Acme"))

    // Add document with explicit topic isolation (atomic: dict + facts + vector + content)
    topicID := uint32(101)
    s.AddDocumentWithTopic(topicID, "auth:Login", sourceCode, embedding, metadata)

    // Cross-subsystem transaction (opt-in, all-or-nothing)
    s.Update(func(txn *meb.StoreTxn) error {
        id, _ := txn.GetOrCreateID("entity:Foo")
        txn.AddFact(meb.Fact{Subject: "entity:Foo", Predicate: "type", Object: "class"})
        txn.AddVector(id, embedding)
        txn.SetContent(id, sourceCode)
        return nil // any error rolls back everything
    })

    // Hybrid search limited to a specific topic
    results, _ := s.Find().
        InTopic(topicID).
        SimilarTo(embedding).
        Limit(5).
        Execute(ctx)

    // Scan (zero-copy streaming, constant memory)
    for f, err := range s.Scan("Alice", "", "") {
        if err != nil {
            log.Fatal(err)
        }
        log.Println(f.String())
    }
}

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                        MEB Store                                 │
├──────────────────────────────────────────────────────────────────┤
│                                                                  │
│  ┌────────────────────────────────────────────────────────────┐ │
│  │  Triple Store          ContentStore      VectorStore       │ │
│  │  (SPO/OPS dual idx)   (S2 blobs)        (Badger-native +  │ │
│  │                                          mmap cache layer) │ │
│  └────────────────────────────────────────────────────────────┘ │
│                                                                  │
│  ┌────────────────────────────────────────────────────────────┐ │
│  │  Query Builder         │  LFTJ Engine                      │ │
│  │  (Neuro-Symbolic)      │  (Streaming joins, visit limits)  │ │
│  └────────────────────────────────────────────────────────────┘ │
│                                                                  │
│  ┌────────────────────────────────────────────────────────────┐ │
│  │  Telemetry (Push)    │  WAL (Crash Recovery)               │ │
│  │  Circuit/GC/Events   │  Single-DB atomicity                │ │
│  └────────────────────────────────────────────────────────────┘ │
│                                                                  │
│  ┌────────────────────────────────────────────────────────────┐ │
│  │  Transaction API (Opt-in)                                   │ │
│  │  View() / Update() — atomic across all subsystems           │ │
│  │  Auto-rollback on error, counter recovery                  │ │
│  └────────────────────────────────────────────────────────────┘ │
│                                                                  │
│  ┌────────────────────────────────────────────────────────────┐ │
│  │  Unified BadgerDB (Graph + Dict + Vectors + Content)       │ │
│  │  Key prefixes: 0x10 content | 0x11 vectors | 0x20/0x21     │ │
│  │  graph | 0x80/0x81 dict | 0xFF system                     │ │
│  └────────────────────────────────────────────────────────────┘ │
│                                                                  │
└──────────────────────────────────────────────────────────────────┘
Hybrid Retrieval Strategy

MEB's Neuro-Symbolic pipeline combines neural vector search with symbolic graph traversal:

1. Semantic Seed (Neuro)
   ┌─────────────────────────────────────────────────────────┐
   │ Dual-path vector search:                                │
   │  • Hot path: Parallel mmap scan in RAM (~500K vectors/s)│
   │  • Cold path: Badger iterator streaming (~50-100K v/s)  │
   │  • Memory: O(k) for top-k heap, never O(N)              │
   │  • Topic search: LSM-level prefix scan skips            │
   │    unrelated topics at storage level                    │
   └────────────────────────┬────────────────────────────────┘
                            │ candidate IDs
                            ▼
2. Structural Expansion (Symbolic)
   ┌─────────────────────────────────────────────────────────┐
   │ LFTJ engine uses seeds to traverse SPO/OPS              │
   │ indices for relational context — streaming, no material │
   └────────────────────────┬────────────────────────────────┘
                            │ filtered triples
                            ▼
3. Lazy Hydration
   ┌─────────────────────────────────────────────────────────┐
   │ Only final filtered results are decompressed from        │
   │ the S2 ContentStore. No unnecessary I/O.                │
   └─────────────────────────────────────────────────────────┘
Key Encoding & ID Structure

MEB uses a Symmetric TopicID packing strategy to enable multi-tenancy within a standard triple-store model — no Graph column needed.

Prefix │ Index │ Key Size │ Value Size │ ID Layout (64-bit)
───────┼───────┼──────────┼────────────┼───────────────────────────────────
0x20   │ SPO   │ 25 bytes │ 16 bytes   │ [Topic:24] | [Local:40]
0x21   │ OPS   │ 25 bytes │ 16 bytes   │ [Topic:24] | [Local:40]
0x10   │ Chunk │ 9 bytes  │ S2 blob    │ [Local:40]
0x11   │ Vec   │ 9 bytes  │ compressed │ [Local:40]
0x80   │ Dict→ │ var      │ 8 bytes    │ string → ID
0x81   │ Dict← │ 9 bytes  │ var        │ ID → string
0xFF   │ Sys   │ 2 bytes  │ counter    │ —

ID Bit-packing:

ID = (TopicID << 40) | LocalID

  TopicID (24-bit): Supports up to 16M isolated namespaces/repositories.
  LocalID (40-bit): Supports over 1 trillion entities per topic.

  ┌──────────────────────┬────────────────────────────────────────────┐
  │   TopicID (24 bits)  │             LocalID (40 bits)              │
  │   bits 63–40         │             bits 39–0                     │
  └──────────────────────┴────────────────────────────────────────────┘

Value Format: [Vector_ID(8) | Content_Offset(8)]

The Content_Offset uses the lower 48 bits, leaving 16 bits for semantic hints (type, hash).

Zero-Copy Streaming

All scan operations leverage Go 1.23+ iter.Seq2 for zero-copy, constant-memory traversal:

// O(1) memory regardless of result set size
// No intermediate slice allocation — results stream directly
for fact, err := range store.Scan("subject", "", "") {
    // Process each fact as it arrives
}
  • Constant Memory: $O(1)$ usage regardless of millions of triples
  • Cloud Run Safe: No OOM on constrained instances
  • Lazy Evaluation: Facts materialized only when consumed
Hybrid Vector Compression

FWHT preconditioning spreads energy uniformly, making block-wise quantization stable across all 1536 dimensions:

Input: 1536-d float32 (OpenAI)
  ↓
Pad to next power of 2 (2048)
  ↓
FWHT (Fast Walsh-Hadamard Transform) — O(N log N), in-place
  ↓
Normalize by 1/√N (unitary)
  ↓
Block-wise Quantize: 64 blocks of 32 elements
  ↓
Per block: [scale:4B][zero:4B][norm:4B][q_0:1B]...[q_31:1B]
  ↓
Compressed: 2,816 bytes (8-bit) or 1,792 bytes (4-bit)

The per-block L2 norm enables Cauchy-Schwarz early termination: suffix sums of normA[r] × normB[r] provide an upper bound on remaining dot product contribution, allowing pruning of dissimilar vectors after scanning only 10-20 of 64 blocks.

BitWidth Compressed Size Compression RAM for 1M vectors
8-bit 2,816 bytes 2.2x 2.82 GB
4-bit 1,792 bytes 3.4x 1.79 GB
float32 6,144 bytes 1x 6.14 GB
Vector Search Architecture

MEB uses a dual-path vector retrieval strategy for optimal performance at any scale:

Search() / SearchWithFilter()
    │
    ├─ mmap has vectors? (totalVectors > 0)
    │   └─ YES → 4-way parallel mmap scan (~500K vectors/sec)
    │            Cache-hot sequential RAM access
    │
    └─ NO → Badger iterator streaming (~50-100K vectors/sec)
            Warms mmap cache as it goes — next search hits fast path

SearchInTopic(topicID)
    └─ Badger prefix scan on [0x11][topicID]
       LSM-level topic filtering — zero I/O on unrelated topics
       Memory: O(k) for top-k heap, never O(N)

Benefits:

  • Hot queries: Parallel mmap scan at ~500K vectors/sec
  • Cold start: Streams from Badger, no RAM limit
  • Topic filtering: Skips 99% of data at storage level
  • Write durability: Synchronous Badger write before return
  • Scale: Limited by disk size, not RAM

Blockwise Dot Product — similarity computed directly on compressed data without full dequantization:

dot(a, b) = Σ_blocks (scale_a * scale_b * Σ(q_a_i * q_b_i)
                      + scale_a * zero_b * Σ(q_a_i)
                      + scale_b * zero_a * Σ(q_b_i)
                      + block_size * zero_a * zero_b)

Pruning — per-block norm suffix sums enable Cauchy-Schwarz early termination:

suffixNormProduct[b] = Σ_{r>=b} normA[r] * queryNorms[r]

for each block b:
    totalSum += blockContribution
    if totalSum + suffixNormProduct[b+1] < threshold:
        return totalSum  // early exit — impossible to beat kth-best

Cross-Subsystem Transactions

MEB provides opt-in transactions for atomic operations across graph, dictionary, vectors, and content. The existing API (AddFact, AddDocument, etc.) works unchanged; transactions are an additional layer for complex use cases.

// Atomic multi-operation: all succeed or all rollback
err := store.Update(func(txn *meb.StoreTxn) error {
    // Dictionary
    id, err := txn.GetOrCreateID("entity:UserService")

    // Graph facts
    txn.AddFact(meb.Fact{Subject: "entity:UserService", Predicate: "type", Object: "class"})
    txn.AddFact(meb.Fact{Subject: "entity:UserService", Predicate: "package", Object: "com.example"})

    // Vector embedding
    txn.AddVector(id, embedding)

    // Content
    txn.SetContent(id, sourceCode)

    return nil // any error → automatic rollback
})

// Read-only transaction
store.View(func(txn *meb.StoreTxn) error {
    for f, err := range txn.Scan(ctx, "entity:UserService", "", "") {
        // ...
    }
    return nil
})

Transaction guarantees:

  • Atomicity: All writes commit together, or none do
  • Rollback: Any error discards all writes in the transaction
  • Counter recovery: numFacts counter is restored on rollback
  • Isolation: Uses BadgerDB's MVCC for read/write isolation

When to use transactions:

Scenario Use AddFact()/AddDocument() Use Update()
Single fact ✅
Single document ✅
Multiple subsystems (dict + facts + vector + content) ✅
Conditional writes ✅
Rollback on error ✅
Hybrid WAL Approach

MEB uses a hybrid WAL strategy to balance durability with performance:

Write Path Durability Mechanism Reason
Transactions (Update()/View()) BadgerDB SyncWrites: true BadgerDB handles transaction durability natively; WAL is redundant
WriteBatch (AddFactBatch()) WAL v2 (CRC32C) BadgerDB WriteBatch lacks SyncWrites; WAL provides crash recovery

Configuration:

  • DefaultConfig() → SyncWrites: true (ingest-heavy, durability-focused)
  • SafeServingConfig() → SyncWrites: false (read-only/serve)
  • ReadOnlyConfig() → SyncWrites: false (read-only)

WAL features:

  • CRC32C checksums for corruption detection
  • replayWAL() for WriteBatch crash recovery (not used for transactions)
  • TruncateIncompleteWAL() for cleaning up partial writes

When WAL is used:

  • WriteBatch path (e.g., bulk ingestion with AddFactBatch)
  • Transactions use BadgerDB's native durability (no WAL)
Background Orphan Cleanup

MEB runs periodic cleanup of orphaned data during runCleanup():

  1. Deprecated triples — Scans SPO index, deletes entries with FlagIsDeprecated
  2. Orphan dictionary entries — Scans reverse dictionary, removes entries not referenced by any fact
  3. Orphan vectors — Two-pass scan: collects all subject IDs from facts, then removes vectors with no content and no referencing facts
  4. ValueLog GC — Runs BadgerDB GC to reclaim space from tombstones

All cleanup operations are throttled (max 1000 orphans per cycle) to avoid long GC pauses.

Package Structure

meb/
├── .github/workflows/
│   └── ci.yml            # CI: go test, goleak, race detector
├── keys/              # 25-byte triple key encoding (TopicID packing)
├── dict/              # String interning (thread-safe LRU + sharded allocator)
├── store/             # BadgerDB config with deployment profiles
├── vector/            # Hybrid vector compression and search
│   ├── turboquant.go  # FWHT + block-wise 4/8-bit quantization + Cauchy-Schwarz pruning
│   ├── partitioned.go # PartitionedRegistry (sharded by TopicID)
│   ├── registry.go    # Badger-native store + mmap cache, RCU revMap
│   ├── search.go      # Dual-path: mmap parallel + Badger streaming, adaptive workers
│   ├── math.go        # L2 normalize, dot product
│   ├── index.go       # Index selection API (brute-force / IVF-PQ / HNSW)
│   ├── ivfpq.go       # IVF-PQ index: centroid assignment, ADC search, posting list iteration
│   ├── ivfpq_config.go# IVFPQConfig with validation
│   ├── ivfpq_train.go # Mini-batch K-means training, PQ codebook training
│   ├── hnsw.go        # HNSW index: Insert, SearchInTopic, searchLayer
│   ├── hnsw_graph.go  # HNSW graph operations: writeNode, readNeighbors, addSymmetricalLink
│   ├── hnsw_delete.go # SoftDelete, isDeleted, Compact
│   ├── adc_scalar.go       # Portable scalar ADC accumulation (4-way unrolled)
│   ├── adc_sse_amd64.go    # SSE ADC kernel for amd64 (dispatch)
│   ├── adc_neon_arm64.go   # NEON ADC kernel for arm64 (dispatch)
│   ├── adc_dispatch_amd64.go   # amd64 runtime dispatch via golang.org/x/sys/cpu
│   ├── adc_dispatch_arm64.go   # arm64 runtime dispatch
│   ├── adc_dispatch_generic.go # Generic fallback dispatch
│   ├── adc_avx2_amd64.s   # Hand-written Plan 9 AVX2 assembly
│   ├── adc_neon_arm64.s   # Hand-written Plan 9 NEON assembly
│   └── diskann.go     # DiskANNConfig and DiskANNIndex scaffold (unreachable)
├── query/             # LFTJ engine + CBO + RBO + plan cache
│   ├── lftj.go        # TrieIterator, LFTJResult, Canonical ordering
│   ├── engine.go      # Execute, ExecuteOrdered, WithBufferAndSort, OptimizeRelations
│   ├── rbo.go         # Rule-Based Optimizer (PlanType, ExecutionPlan, Optimize)
│   ├── rbo_test.go    # RBO unit tests
│   ├── plan_cache.go  # QueryPlanCache with LRU eviction + TTL + FNV-1a hash
│   └── stream.go      # ExecuteStream concurrent dense/sparse merge
├── circuit/           # Query timeout circuit breaker with state callbacks
├── utils/             # Zero-copy string/byte conversion
├── adapter/           # Mangle Datalog integration
├── bench/             # ANN benchmarks + perf benchmarks
├── store.go           # MEBStore orchestrator (NewMEBStore, Reset, Close, EnableIVFPQ, EnableHNSW)
├── tx.go              # Transaction API (View/Update, StoreTxn, AddIVFVector, SearchHybrid)
├── knowledge_store.go # SPO/OPS dual-index write, orphan cleanup
├── scan.go            # Index selection scan (iter.Seq2 streaming)
├── content.go         # S2-compressed content storage, atomic Add/DeleteDocument
├── query_builder.go   # Neuro-symbolic query builder with LFTJ joins — Execute(ctx)
├── telemetry.go       # Push telemetry (TelemetrySink interface)
├── wal.go             # Write-ahead log v2 (CRC32C, WriteBatch crash recovery only)
├── fact_store.go      # factstore.FactStore implementation
├── store_test.go      # H4 Lifecycle, H5 cancellation, H10 lifecycle
├── reset_test.go      # H1 Reset no-deadlock, concurrent guards
├── scan_test.go       # H9 PreserveObjectTypes deprecation warning
├── wal_test.go        # H8 WAL v2 format, CRC32C, WriteBatch recovery tests
└── go.mod             # go.uber.org/goleak added for test goroutine leak detection

Configuration

// Default (Ingest-Heavy)
cfg := store.DefaultConfig("./data")
// SyncWrites: true (transactions use BadgerDB durability)

// Cloud Run optimized
cfg := store.SafeServingConfig("./data")
// SyncWrites: false (read-only/serve mode)

// Read-only
cfg := store.ReadOnlyConfig("./data")
// SyncWrites: false (read-only mode)

// Enable verbose debug logging
cfg.Verbose = true

// Enable sharded dictionary (must be power of 2)
cfg.NumDictShards = 4

// Set vector default dimensionality (default: 1536)
cfg.VectorFullDim = 1536

Observability

MEB provides push telemetry for autonomous agent consumption:

// Register a telemetry sink
store.RegisterTelemetrySink(&mySink{})

// Events emitted automatically:
//   "circuit_state_change" — circuit breaker transitions
//   "gc_failure" — ValueLogGC errors
//   "retention" — fact count exceeds threshold
//   "deprecated_cleanup" — deprecated triples purged
//   "dict_orphan_cleanup" — orphaned dictionary entries removed
//   "vector_orphan_cleanup" — orphaned vectors removed
//   "wal_clear_failed" — WAL consistency issue

Performance

Benchmark results (AMD Ryzen 9 5900HS with Radeon Graphics, in-memory BadgerDB, Go 1.23):

Benchmark Ops/sec Latency Memory Allocs
Fact Insertion (single) 33,507 35.0 µs/op 8.3 KB/op 192
Fact Insertion (batch × 10) 9,794 134.6 µs/op 45.2 KB/op 1,042
Fact Insertion (batch × 100) 1,291 1,094 µs/op 424 KB/op 9,503
Fact Insertion (batch × 1000) 100 10.6 ms/op 4.7 MB/op 93,330
Document Add (content + vector + metadata) 14,925 76.1 µs/op 33.9 KB/op 113
Transaction Batch (100 facts) 2,232 582 µs/op 257 KB/op 5,973
Scan (1000 facts, single key) 21,270 54.9 µs/op 27.2 KB/op 621
Vector Add (1536-d, 8-bit Hybrid) 24,759 47.3 µs/op 29.9 KB/op 37
Vector Search (10K vectors, k=10, with pruning) 273 4.5 ms/op 17.5 KB/op 52
Dictionary Lookup (GetOrCreate) 1,532,017 839 ns/op 400 B/op 7

Derived throughput:

Metric Value Notes
Vector Ingestion ~25K vectors/sec Includes FWHT + quantization + norm computation + Badger write
Vector Search ~2.7M vectors/sec Pruned mmap scan over 10K vectors, 15% faster than pre-pruning
Fact Ingestion ~34K facts/sec (single), ~129K facts/sec (batch × 100) Dual-index SPO+OPS write
Scan Throughput ~21M keys/sec Key-only, SPO prefix scan
Dictionary Lookup ~1.5M lookups/sec Sharded LRU cache hit

Key observations:

  • Cauchy-Schwarz pruning speeds up search by 15% at 10K scale (238→273 ops/s), with increasing gains at larger vector counts
  • Recall@10 improved from 90.0% to 97.0% with per-block norm storage
  • Vector Add is 1.9x faster than original after eliminating double quantization (previously 15K→28K vectors/sec)
  • Vector search latency scales with dimension — 1536-dim is ~40x slower than 128-dim (4.5 ms vs 124 µs)
  • Scan latency scales with matching facts, not total graph size (prefix scan)
  • Dictionary lookups are sub-microsecond with thread-safe LRU cache
  • DeleteFactsBySubject uses scoped prefix scans (not full graph scan) for orphan cleanup
  • Storage overhead: +10% per vector (2,560→2,816 bytes) for per-block norm pruning

Storage scaling (per 1M items):

Mode Storage Components Used
Facts only ~150 MB SPO(41) + OPS(41) + Dict(23) per fact
+ 100K vectors (8-bit) ~430 MB Facts(150) + Vectors(282)
+ 100K vectors + 10K docs (S2) ~530 MB Facts(150) + Vectors(282) + Content(100)
  • Facts-only mode uses minimal storage — no overhead from unused subsystems
  • Vectors dominate storage (4-5x larger than facts at 8-bit)
  • Content is moderate if documents are few (S2 compressed)
Metric Value Notes
RAM Density Up to 1.1M nodes (1536-d) Within 2GB RAM using Hybrid 4-bit
Disk-Scaled Unlimited vectors Badger-backed storage — not RAM-limited
Cold Start < 200ms warm-up Safe-Serving profile
Join Latency Sub-2s Complex code-graph traversals with circuit breaker
Vector Search (hot, pruned) ~273K vectors/sec mmap parallel scan with Cauchy-Schwarz pruning
Vector Search (cold) ~76M vectors/sec Badger iterator streaming
Topic Search LSM prefix scan only Zero I/O on unrelated topics
Fact Insertion ~102K facts/sec (single), ~861K facts/sec (batch) Dual-index write
Scan Throughput ~102M keys/sec Key-only, SPO prefix scan
Content Read ~500MB/s S2 decompression

Fidelity Verification

Five stress-tested metrics validate the Hybrid (FWHT + Block-wise) quantization under Gaussian distribution with extreme outliers (50x-100x spikes):

Metric Test Result Target Status
Mathematical Fidelity TestHybridQuantizationFidelity 8-bit: 1.0000, 4-bit: 0.9977 8-bit > 0.95, 4-bit > 0.90 ✅
Dot Product Accuracy TestHybridQuantizationFidelity 8-bit MAE: 0.18, 4-bit MAE: 2.97 Rank-order stability ✅
Recall@10 TestRecallAtK (10K vectors, 10 seeds) 97.0% avg > 80% ✅
FWHT Invariance TestFWHTInvariance Max error < 1e-5 (dims 4-2048) FWHT(FWHT(v))/N == v ✅
Energy Spreading TestQuantizationDistribution CV: 0.29 → 0.16 (46% reduction) avgCVFWHT < avgCVNoFWHT ✅
8-bit Lossless Test8BitLosslessVerification Max error: 0.063, BER: 29.8% Max error < 1.0, BER < 50% ✅

Key observations:

  • 8-bit is near-lossless (cosine = 1.0000) — 256 levels are more than enough for 32-element blocks
  • 4-bit fidelity is excellent (cosine = 0.9977) — far exceeds the 0.90 target even with spikes
  • Recall@10 = 97.0% — improved from 90.0% with per-block norm storage in quantized vectors
  • FWHT reduces block scale variance by 46% (CV: 0.29 → 0.16) — proves energy spreading works on high-entropy data
  • Dot product MAE stays low — 8-bit MAE of 0.18 vs 4-bit MAE of 2.97 on 1536-dim vectors
  • FWHT is mathematically correct — passes invariance test across all power-of-2 dimensions

Run fidelity tests:

go test ./vector/... -v -run "TestHybridQuantizationFidelity|TestRecallAtK|TestFWHTInvariance|TestQuantizationDistribution|Test8BitLosslessVerification"

Cloud Run Guardrails:

Constraint Value Purpose
Query Circuit Breaker 2,000ms Stop runaway queries
Max Join Results 5,000 facts Prevent RAM exhaustion
Vector Search Top-K 100 Limit result set size

Build

go build ./...
go test -short ./...    # skip slow ANN benchmarks
go test -race ./...     # race detector (except vector/ — FWHT timeout)
go vet ./...

Dependencies

  • github.com/dgraph-io/badger/v4 — Key-value storage
  • codeberg.org/TauCeti/mangle-go — Datalog reasoning engine
  • github.com/klauspost/compress — S2 compression
  • github.com/hashicorp/golang-lru/v2 — Dictionary caching
  • go.uber.org/goleak — Goroutine leak detection in tests
  • golang.org/x/sys — CPU feature detection (AVX2 dispatch)

Documentation

Index

Constants

View Source 
const (
	DefaultCandidateMultiplier = 10
)

Variables

View Source 
var (
	// ErrInvalidFact is returned when a Fact has missing required fields
	// (empty Subject or Predicate) or an Object of an unsupported type.
	ErrInvalidFact = fmt.Errorf("invalid fact")

	// ErrStoreReadOnly is returned when a write operation is attempted on a
	// store opened with Config.ReadOnly = true.
	ErrStoreReadOnly = fmt.Errorf("store is read-only")

	// ErrInvalidQuery is returned when a query is malformed (e.g. empty
	// predicate in a context that requires one).
	ErrInvalidQuery = fmt.Errorf("invalid query")

	// ErrFactNotFound is returned by lookup helpers when no matching fact
	// is found.
	ErrFactNotFound = fmt.Errorf("fact not found")

	// ErrEmptyBatch is returned when a batch write is called with no facts.
	ErrEmptyBatch = fmt.Errorf("empty batch")

	// ErrStoreClosed is returned when an operation is attempted on a store or
	// batch writer that has been closed.
	ErrStoreClosed = fmt.Errorf("store is closed")

	// ErrWALClosed is returned by WAL.Append or WAL.Clear when the WAL has
	// been closed or is mid-clear. Callers should treat this as a hard
	// failure rather than a silent drop.
	ErrWALClosed = fmt.Errorf("WAL is closed")

	// ErrUnknownTriplePrefix is returned by key encoders when an unknown
	// prefix byte is passed. Surfaces bugs that would otherwise be silently
	// mis-encoded as SPO.
	ErrUnknownTriplePrefix = fmt.Errorf("unknown triple key prefix")
)

Functions

func CountFacts 

func CountFacts(seq iter.Seq2[Fact, error]) (int, error)

func Filter 

func Filter(seq iter.Seq2[Fact, error], pred func(Fact) bool) iter.Seq2[Fact, error]

func Map 

func Map[T any](seq iter.Seq2[Fact, error], fn func(Fact) (T, error)) iter.Seq2[T, error]

func MustValue 

func MustValue[T any](f Fact) T

func ShouldPruneTriple 

func ShouldPruneTriple(value []byte, wantEntityType uint16, wantPublic bool) bool

ShouldPruneTriple reads semantic hints from a 16-byte triple value and returns true if the triple should be pruned based on the requested entity type and public flag.

func Value 

func Value[T any](f Fact) (T, bool)

func ValueOrDefault 

func ValueOrDefault[T any](f Fact, defaultVal T) T

Types

type BatchWriter added in v0.7.0 

type BatchWriter struct {
	// contains filtered or unexported fields
}

BatchWriter accumulates bulk triple writes and commits them through a single BadgerDB WriteBatch, amortizing fsync/commit overhead across many facts. It is intended for bulk/A^-style imports. Facts must be added via AddFacts; the batch is committed with Flush (or discarded with Cancel). A BatchWriter is single-goroutine and must not be reused after Flush or Cancel.

func (*BatchWriter) AddFacts added in v0.7.0 

func (bw *BatchWriter) AddFacts(facts []Fact) error

AddFacts queues the given facts for the batch. Facts whose triple already exists in the store (or was added earlier in this batch) are skipped and do not affect Flush's count, matching the transaction path's numFacts semantics.

func (*BatchWriter) Cancel added in v0.7.0 

func (bw *BatchWriter) Cancel()

Cancel discards the batch without committing, releasing the WriteBatch.

func (*BatchWriter) Count added in v0.7.0 

func (bw *BatchWriter) Count() uint64

Count returns the number of new facts queued in this batch so far (i.e. those that will be reflected in the store's fact count after Flush).

func (*BatchWriter) Flush added in v0.7.0 

func (bw *BatchWriter) Flush() error

Flush commits the buffered writes, updates the store fact count with the number of newly inserted triples, and closes the batch. Flush may be called only once. Like the transaction path, it relies on Badger durability and does not write the WAL (see BatchWriter doc on AddFacts).

type Builder 

type Builder struct {
	// contains filtered or unexported fields
}

func NewBuilder 

func NewBuilder(store Store) *Builder

func (*Builder) CandidateMultiplier 

func (b *Builder) CandidateMultiplier(multiplier int) *Builder

func (*Builder) Execute 

func (b *Builder) Execute(ctx context.Context) ([]Result, error)

Execute runs the query using RBO to automatically select the optimal strategy. FilterFirst() is kept as an explicit override for manual control.

func (*Builder) FilterFirst 

func (b *Builder) FilterFirst() *Builder

FilterFirst enables candidate-set pre-filtering for this query. When enabled, filters are evaluated against the graph index FIRST, building a set of matching subject IDs. Vector search results that aren't in this set are then skipped in the result loop (post-filter).

Important: This is NOT true predicate pushdown — vector search still scans all vectors and computes similarities for all candidates. The benefit comes from skipping dictionary lookups and content fetches for non-matching vectors. Most effective when:

  • Filters are highly selective (< 10% match rate)
  • Content fetch or dict lookup is expensive relative to vector search

func (*Builder) InTopic 

func (b *Builder) InTopic(topicID uint32) *Builder

InTopic restricts the search to a specific topic. The TopicID is used for topic-aware vector search and scan operations.

func (*Builder) JoinWithLFTJ 

func (b *Builder) JoinWithLFTJ(relations []query.RelationPattern, resultVars []string) *Builder

JoinWithLFTJ adds LFTJ relations for structural expansion after vector search. The seedVar must match a variable name used in the relations. Results are streamed — no intermediate materialization.

func (*Builder) Limit 

func (b *Builder) Limit(n int) *Builder

func (*Builder) SimilarTo 

func (b *Builder) SimilarTo(vec []float32) *Builder

func (*Builder) SimilarToWithThreshold 

func (b *Builder) SimilarToWithThreshold(vec []float32, threshold float32) *Builder

func (*Builder) Where 

func (b *Builder) Where(predicate string, object interface{}) *Builder

type Fact 

type Fact struct {
	Subject   string
	Predicate string
	Object    any
}

func Collect 

func Collect(seq iter.Seq2[Fact, error]) ([]Fact, error)

func First 

func First(seq iter.Seq2[Fact, error]) (Fact, error)

func NewFact 

func NewFact(subject, predicate string, object any) Fact

func (Fact) IsValid 

func (f Fact) IsValid() bool

func (Fact) String 

func (f Fact) String() string

type FilterOpt 

type FilterOpt struct {
	Predicate string
	Object    interface{}
}

FilterOpt is a filter option for hybrid search.

type MEBStore 

type MEBStore struct {
	// contains filtered or unexported fields
}

func NewMEBStore 

func NewMEBStore(cfg *store.Config) (*MEBStore, error)

func (*MEBStore) Add 

func (m *MEBStore) Add(atom ast.Atom) bool

func (*MEBStore) AddDocument 

func (m *MEBStore) AddDocument(docKey string, content []byte, vec []float32, metadata map[string]any) error

func (*MEBStore) AddDocumentWithTopic 

func (m *MEBStore) AddDocumentWithTopic(topicID uint32, docKey string, content []byte, vec []float32, metadata map[string]any) error

func (*MEBStore) AddFact 

func (m *MEBStore) AddFact(fact Fact) error

func (*MEBStore) AddFactBatch 

func (m *MEBStore) AddFactBatch(facts []Fact) error

func (*MEBStore) CircuitBreaker 

func (m *MEBStore) CircuitBreaker() *circuit.Breaker

func (*MEBStore) CircuitBreakerMetrics 

func (m *MEBStore) CircuitBreakerMetrics() circuit.Metrics

func (*MEBStore) CircuitBreakerMetricsSnapshot 

func (m *MEBStore) CircuitBreakerMetricsSnapshot() circuit.MetricsSnapshot

func (*MEBStore) Close 

func (m *MEBStore) Close() error

func (*MEBStore) Compact added in v0.7.0 

func (m *MEBStore) Compact(ratio float64) error

Compact runs Badger value-log garbage collection, rewriting stale values to reclaim disk space. ratio is the fraction of a value log file's bytes that must be discardable to justify rewriting it (0-1; badger.ErrNoRewrite is returned when there is nothing to reclaim). The store's configured GCRatio is used if ratio is <= 0.

func (*MEBStore) Contains 

func (m *MEBStore) Contains(atom ast.Atom) bool

func (*MEBStore) Count 

func (m *MEBStore) Count() uint64

func (*MEBStore) DebugInfo added in v0.6.0 

func (m *MEBStore) DebugInfo() StoreHealth

DebugInfo returns a snapshot of store health metrics. Safe to call concurrently; all reads are atomic or lock-protected.

func (*MEBStore) DeleteDocument 

func (m *MEBStore) DeleteDocument(docKey string) error

func (*MEBStore) DeleteDocumentWithTopic 

func (m *MEBStore) DeleteDocumentWithTopic(docKey string, topicID uint32) error

func (*MEBStore) DeleteFactsBySubject 

func (m *MEBStore) DeleteFactsBySubject(subject string) error

func (*MEBStore) Dict 

func (m *MEBStore) Dict() dict.Dictionary

func (*MEBStore) EnableHNSW 

func (st *MEBStore) EnableHNSW(cfg *vector.HNSWConfig) error

func (*MEBStore) EnableIVFPQ 

func (st *MEBStore) EnableIVFPQ(cfg *vector.IVFPQConfig) error

func (*MEBStore) Exists 

func (m *MEBStore) Exists(s, p, o string) bool

Exists performs an efficient key-only existence check without decoding strings.

func (*MEBStore) Find 

func (m *MEBStore) Find() *Builder

func (*MEBStore) FindSubjectsByObject 

func (m *MEBStore) FindSubjectsByObject(ctx context.Context, predicate, object string) iter.Seq[string]

FindSubjectsByObject returns all subjects matching exact predicate and object. Uses SPO index scan across ALL topics - does not filter by current topicID. Returns empty iterator if no matches found (never returns error for "not found").

func (*MEBStore) GetContent 

func (m *MEBStore) GetContent(id uint64) ([]byte, error)

func (*MEBStore) GetContentByKey 

func (m *MEBStore) GetContentByKey(docKey string) ([]byte, error)

func (*MEBStore) GetDocumentMetadata 

func (m *MEBStore) GetDocumentMetadata(docKey string) (map[string]any, error)

func (*MEBStore) GetFacts 

func (m *MEBStore) GetFacts(atom ast.Atom, callback func(ast.Atom) error) error

func (*MEBStore) GetTriples added in v0.7.0 

func (m *MEBStore) GetTriples(triples []TripleKey) ([]TripleResult, error)

GetTriples performs a batched point lookup for the given triple keys, resolving all dictionary IDs within a single read transaction. The result slice has the same length as keys, in the same order; Found is false for any key that does not resolve (unknown subject/predicate/object) or has no matching fact.

func (*MEBStore) GetVectors added in v0.7.0 

func (m *MEBStore) GetVectors(ids []uint64) ([]VectorResult, error)

GetVectors performs a batched read of the stored vectors for the given IDs, dequantizing each within a single read transaction. Missing IDs are reported with Found=false.

func (*MEBStore) HNSWIndex added in v0.6.1 

func (st *MEBStore) HNSWIndex() *vector.HNSWIndex

func (*MEBStore) HasDocument 

func (m *MEBStore) HasDocument(docKey string) (bool, error)

func (*MEBStore) Health added in v0.7.0 

func (m *MEBStore) Health() StoreHealth

Health returns a snapshot of store health metrics. It is a convenience alias for DebugInfo, suited to monitoring/health endpoints. Safe to call at any time (including after Close); values simply reflect the closed state.

func (*MEBStore) IVFPQIndex 

func (st *MEBStore) IVFPQIndex() *vector.IVFPQIndex

func (*MEBStore) LFTJEngine 

func (m *MEBStore) LFTJEngine() *query.LFTJEngine

func (*MEBStore) ListPredicates 

func (m *MEBStore) ListPredicates() []ast.PredicateSym

func (*MEBStore) LookupID 

func (m *MEBStore) LookupID(key string) (uint64, bool)

func (*MEBStore) Merge 

func (m *MEBStore) Merge(other factstore.ReadOnlyFactStore) error

func (*MEBStore) MergeBatch 

func (m *MEBStore) MergeBatch(other factstore.ReadOnlyFactStore, batchSize int) error

MergeBatch imports facts from another store in batches, avoiding the O(N²) per-fact Exists() check. Existing facts are silently skipped.

func (*MEBStore) NewBatchWriter added in v0.7.0 

func (m *MEBStore) NewBatchWriter() *BatchWriter

NewBatchWriter opens a new bulk writer against the store's default topic. Callers must either Flush or Cancel to release the underlying WriteBatch.

func (*MEBStore) RecalculateStats 

func (m *MEBStore) RecalculateStats() (uint64, error)

func (*MEBStore) RegisterTelemetrySink 

func (m *MEBStore) RegisterTelemetrySink(sink TelemetrySink)

func (*MEBStore) Reset 

func (m *MEBStore) Reset() error

func (*MEBStore) ResolveID 

func (m *MEBStore) ResolveID(id uint64) (string, error)

func (*MEBStore) RunValueLogGC 

func (m *MEBStore) RunValueLogGC(ratio float64) error

func (*MEBStore) Scan 

func (m *MEBStore) Scan(s, p, o string) iter.Seq2[Fact, error]

func (*MEBStore) ScanContext 

func (m *MEBStore) ScanContext(ctx context.Context, s, p, o string) iter.Seq2[Fact, error]

func (*MEBStore) ScanInTopic 

func (m *MEBStore) ScanInTopic(topicID uint32, s, p, o string) iter.Seq2[Fact, error]

ScanInTopic scans facts within a specific topic. The TopicID is packed into the ID structure for data locality. This enables scanning only the prefix range belonging to the requested topic.

func (*MEBStore) ScanInTopicContext 

func (m *MEBStore) ScanInTopicContext(ctx context.Context, topicID uint32, s, p, o string) iter.Seq2[Fact, error]

ScanInTopicContext scans facts within a specific topic with context.

func (*MEBStore) ScanSubjects 

func (m *MEBStore) ScanSubjects(ctx context.Context) iter.Seq[string]

ScanSubjectsByPrefix returns all subjects starting with the given prefix string. Uses SPO index with LSM-tree prefix scan - O(log N + k) where k is number of results. The prefix is matched against the full subject string (e.g., "project/pkg/" matches all subjects under that path like "project/pkg/server/server.go:NewServer"). Returns empty iterator if no matches found (never returns error for "not found"). ScanSubjects returns all subjects in the store by scanning the SPO index. Warning: This performs a full table scan.

func (*MEBStore) ScanSubjectsByPrefix 

func (m *MEBStore) ScanSubjectsByPrefix(ctx context.Context, prefix string) iter.Seq[string]

func (*MEBStore) ScanWithFilters 

func (m *MEBStore) ScanWithFilters(s, p, o string, filters []PredicateFilter) iter.Seq2[Fact, error]

func (*MEBStore) ScanWithFiltersContext 

func (m *MEBStore) ScanWithFiltersContext(ctx context.Context, s, p, o string, filters []PredicateFilter) iter.Seq2[Fact, error]

func (*MEBStore) ScanWithPruning 

func (m *MEBStore) ScanWithPruning(ctx context.Context, s, p, o string, entityType uint16, wantPublic bool) iter.Seq2[Fact, error]

ScanWithPruning scans facts with semantic hints pruning. entityType: only yield triples matching this entity type (0 = no pruning). wantPublic: if true, only yield triples with IsPublic flag.

func (*MEBStore) SetCircuitBreakerConfig 

func (m *MEBStore) SetCircuitBreakerConfig(config *circuit.Config)

func (*MEBStore) SetContent 

func (m *MEBStore) SetContent(id uint64, data []byte) error

func (*MEBStore) SetDefaultEntityType 

func (m *MEBStore) SetDefaultEntityType(entityType uint16)

SetDefaultEntityType sets the default entity type for semantic hints. Use keys.EntityFunc, EntityVar, EntityClass, etc.

func (*MEBStore) SetDefaultFlags 

func (m *MEBStore) SetDefaultFlags(flags uint16)

SetDefaultFlags sets the default flags for semantic hints. Use keys.FlagIsPublic, FlagIsDeprecated, FlagIsTest, FlagIsGenerated.

func (*MEBStore) SetRetention 

func (m *MEBStore) SetRetention(maxFacts uint64) error

func (*MEBStore) SetTopicID 

func (m *MEBStore) SetTopicID(topicID uint32)

SetTopicID sets the 24-bit topic ID for symmetric bit-packing. All facts added after this call will use the new topic ID. Supports up to 16M isolated namespaces (topics). Panics if topicID is 0 (reserved as invalid).

func (*MEBStore) TopicID 

func (m *MEBStore) TopicID() uint32

TopicID returns the current topic ID.

func (*MEBStore) TrainIVFPQ 

func (st *MEBStore) TrainIVFPQ(topicID uint32) error

func (*MEBStore) UnregisterTelemetrySink 

func (m *MEBStore) UnregisterTelemetrySink(sink TelemetrySink)

func (*MEBStore) Update 

func (m *MEBStore) Update(fn func(*StoreTxn) error) error

Update executes a read-write transaction with automatic commit/rollback. In-memory side effects (counters, mmap caches, dict caches) are applied only after the Badger transaction commits successfully, using accumulated deltas. BadgerDB SyncWrites: true ensures durability for the transaction path. If a post-commit side effect fails the committed transaction is NOT rolled back; the error is returned so callers can detect the inconsistency.

func (*MEBStore) Vectors 

func (m *MEBStore) Vectors() *vector.VectorRegistry

func (*MEBStore) View 

func (m *MEBStore) View(fn func(*StoreTxn) error) error

View executes a read-only transaction.

type NumericRange 

type NumericRange struct {
	Min float64
	Max float64
}

NumericRange represents a closed [Min, Max] numeric range for PredicateRange filters.

type PredicateFilter 

type PredicateFilter struct {
	Type  PredicateFilterType
	Value interface{} // string for regex/contains; float64 for gt/lt/gte/lte; NumericRange for range
	// contains filtered or unexported fields
}

func MustPredicateFilter 

func MustPredicateFilter(filterType PredicateFilterType, value interface{}) *PredicateFilter

MustPredicateFilter creates a PredicateFilter, panicking on invalid input.

func NewPredicateFilter 

func NewPredicateFilter(filterType PredicateFilterType, value interface{}) (*PredicateFilter, error)

NewPredicateFilter creates a PredicateFilter, pre-compiling regex patterns.

type PredicateFilterType 

type PredicateFilterType string
const (
	PredicateRegex    PredicateFilterType = "regex"
	PredicateRange    PredicateFilterType = "range"
	PredicateGT       PredicateFilterType = "gt"
	PredicateLT       PredicateFilterType = "lt"
	PredicateGTE      PredicateFilterType = "gte"
	PredicateLTE      PredicateFilterType = "lte"
	PredicateContains PredicateFilterType = "contains"
)

type QueryFilter 

type QueryFilter struct {
	Predicate string
	Object    interface{}
}

type Result 

type Result struct {
	ID      uint64  // Internal dictionary ID
	Key     string  // Human-readable key (decoded from dictionary)
	Score   float32 // Similarity score (0-1, higher is better)
	Content string  // Document content (empty if not found)
}

Result represents a single neuro-symbolic search result.

func (Result) String 

func (r Result) String() string

String returns a string representation of the result.

type Results 

type Results []Result

Results is a slice of Result with helper methods.

func (Results) IDs 

func (rs Results) IDs() []uint64

IDs returns all result IDs.

func (Results) Keys 

func (rs Results) Keys() []string

Keys returns all result keys.

func (Results) Scores 

func (rs Results) Scores() []float32

Scores returns all result scores.

type Store 

type Store interface {
	Vectors() *vector.VectorRegistry
	Scan(s, p, o string) iter.Seq2[Fact, error]
	GetContent(id uint64) ([]byte, error)
	ResolveID(id uint64) (string, error)
	LFTJEngine() *query.LFTJEngine
	Dict() dict.Dictionary
	Exists(s, p, o string) bool
}

type StoreHealth added in v0.6.0 

type StoreHealth struct {
	NumFacts              uint64                  `json:"num_facts"`
	NumVectors            int64                   `json:"num_vectors"`
	VectorFullDim         int                     `json:"vector_full_dim"`
	VectorCapacity        int                     `json:"vector_capacity"`
	WALSizeBytes          int64                   `json:"wal_size_bytes,omitempty"`
	ReadOnly              bool                    `json:"read_only"`
	CircuitBreakerState   string                  `json:"circuit_breaker_state"`
	CircuitBreakerMetrics circuit.MetricsSnapshot `json:"circuit_breaker_metrics,omitempty"`
	LastGCTimeNano        int64                   `json:"last_gc_time_nano"`
	StoreOpen             bool                    `json:"store_open"`
	LSMSize               int64                   `json:"lsm_size_bytes,omitempty"`
	ValueLogSize          int64                   `json:"value_log_size_bytes,omitempty"`
	DiskUsage             int64                   `json:"disk_usage_bytes,omitempty"`
	MemoryInUse           uint64                  `json:"memory_in_use_bytes,omitempty"`
}

StoreHealth provides operational metrics for debugging and monitoring. All values are atomically or lock-protected reads; safe to call concurrently.

type StoreTxn 

type StoreTxn struct {
	// contains filtered or unexported fields
}

StoreTxn wraps a BadgerDB transaction with store-specific helpers. Created via MEBStore.View() or MEBStore.Update().

func (*StoreTxn) AddFact 

func (t *StoreTxn) AddFact(fact Fact) error

AddFact adds a single fact within the transaction.

func (*StoreTxn) AddFactBatch 

func (t *StoreTxn) AddFactBatch(facts []Fact) error

AddFactBatch adds multiple facts within the transaction. Uses the store's current topicID. Durability is provided by BadgerDB SyncWrites: true (no WAL needed for txn path).

func (*StoreTxn) AddFactBatchWithTopic 

func (t *StoreTxn) AddFactBatchWithTopic(facts []Fact, topicID uint32) error

AddFactBatchWithTopic adds multiple facts within the transaction using a specific topicID. Durability is provided by BadgerDB SyncWrites: true (no WAL needed for txn path).

func (*StoreTxn) AddHNSWVector 

func (t *StoreTxn) AddHNSWVector(topicID uint32, localID uint64, fullVec []float32) error

AddHNSWVector adds a vector to the HNSW index.

func (*StoreTxn) AddIVFVector 

func (t *StoreTxn) AddIVFVector(topicID uint32, localID uint64, fullVec []float32) error

AddIVFVector adds a vector to the IVF-PQ index.

func (*StoreTxn) AddVector 

func (t *StoreTxn) AddVector(id uint64, vec []float32) error

AddVector adds a vector within the transaction. The Badger write is done in-txn; the mmap cache update is deferred to post-commit.

func (*StoreTxn) AddVectorWithHash 

func (t *StoreTxn) AddVectorWithHash(id uint64, vec []float32, semanticHash uint8) error

AddVectorWithHash adds a vector with a semantic hash within the transaction.

func (*StoreTxn) BadgerTxn 

func (t *StoreTxn) BadgerTxn() *badger.Txn

BadgerTxn returns the underlying Badger transaction (for debugging).

func (*StoreTxn) DeleteDocument 

func (t *StoreTxn) DeleteDocument(docKey string) error

DeleteDocument atomically deletes a document (content, vector, facts, dict entry) within the transaction. Uses the store's current topicID. In-memory vector/dict mutations are deferred to post-commit.

func (*StoreTxn) DeleteDocumentWithTopic 

func (t *StoreTxn) DeleteDocumentWithTopic(docKey string, topicID uint32) error

DeleteDocumentWithTopic atomically deletes a document using a specific topicID.

func (*StoreTxn) DeleteFactsBySubject 

func (t *StoreTxn) DeleteFactsBySubject(subject string) error

DeleteFactsBySubject deletes all facts for the given subject within the transaction. Uses the store's current topicID.

func (*StoreTxn) DeleteVector 

func (t *StoreTxn) DeleteVector(id uint64) bool

DeleteVector deletes a vector within the transaction. In-memory registry mutation is deferred to post-commit.

func (*StoreTxn) Exists 

func (t *StoreTxn) Exists(s, p, o string) bool

Exists checks if a fact exists within the transaction.

func (*StoreTxn) GetContent 

func (t *StoreTxn) GetContent(id uint64) ([]byte, error)

GetContent retrieves content by ID within the transaction.

func (*StoreTxn) GetID 

func (t *StoreTxn) GetID(s string) (uint64, error)

GetID gets a dictionary ID for the given string within the transaction.

func (*StoreTxn) GetOrCreateID 

func (t *StoreTxn) GetOrCreateID(s string) (uint64, error)

GetOrCreateID gets or creates a dictionary ID for the given string within the transaction.

func (*StoreTxn) GetString 

func (t *StoreTxn) GetString(id uint64) (string, error)

GetString resolves a dictionary ID to its string within the transaction.

func (*StoreTxn) HasVector 

func (t *StoreTxn) HasVector(id uint64) bool

HasVector checks if a vector exists within the transaction.

func (*StoreTxn) Scan 

func (t *StoreTxn) Scan(ctx context.Context, s, p, o string) iterSeq2FactError

Scan iterates over facts matching the given subject, predicate, object. Uses the store's current topicID.

func (*StoreTxn) ScanInTopic 

func (t *StoreTxn) ScanInTopic(ctx context.Context, topicID uint32, s, p, o string) iterSeq2FactError

ScanInTopic iterates over facts matching the given subject, predicate, object within the specified topic.

func (*StoreTxn) SearchHNSW 

func (t *StoreTxn) SearchHNSW(ctx context.Context, topicID uint32, queryVec []float32, k int) ([]Result, error)

SearchHNSW searches the HNSW index directly (bypassing RBO).

func (*StoreTxn) SearchHybrid 

func (t *StoreTxn) SearchHybrid(ctx context.Context, queryVec []float32, k int) ([]Result, error)

SearchHybrid uses the RBO to select the optimal search strategy.

func (*StoreTxn) SearchHybridWithFilters 

func (t *StoreTxn) SearchHybridWithFilters(ctx context.Context, queryVec []float32, k int, filters []FilterOpt) ([]Result, error)

SearchHybridWithFilters is like SearchHybrid but accepts optional FilterOpt filters.

func (*StoreTxn) SearchIVFPQ 

func (t *StoreTxn) SearchIVFPQ(ctx context.Context, topicID uint32, queryVec []float32, k int) ([]Result, error)

SearchIVFPQ searches the IVF-PQ index directly (bypassing RBO).

func (*StoreTxn) SetContent 

func (t *StoreTxn) SetContent(id uint64, data []byte) error

SetContent stores content by ID within the transaction.

type TelemetryEvent 

type TelemetryEvent struct {
	Type      string
	Timestamp time.Time
	Data      map[string]any
}

type TelemetrySink 

type TelemetrySink interface {
	OnEvent(event TelemetryEvent)
}

type TripleKey added in v0.7.0 

type TripleKey struct {
	Subject   string
	Predicate string
	Object    any
}

TripleKey identifies a triple for point lookup. Unlike a scan (which matches by prefix/wildcard), every component must be fully bound. The Object may be any value the store can encode (string, bool, int32, float32, or a dictionary-backed scalar).

type TripleResult added in v0.7.0 

type TripleResult struct {
	Key   TripleKey
	Found bool
	Fact  Fact
}

TripleResult carries the outcome of a single GetTriples lookup.

type VectorResult added in v0.7.0 

type VectorResult struct {
	ID    uint64
	Found bool
	Vec   []float32
}

VectorResult carries the outcome of a single GetVectors lookup. The vector is the lossy dequantized reconstruction of the stored block-quantized hybrid data (matches what search operates on), not the original input vector.

type WAL 

type WAL struct {
	// contains filtered or unexported fields
}

func NewWAL 

func NewWAL(dataDir string) (*WAL, error)

NewWAL opens or creates a v2 WAL file. Returns an error if an existing WAL file has an unsupported format (e.g. v1) — the caller should delete the file.

func (*WAL) Append 

func (w *WAL) Append(entry walEntry) error

Append writes a single entry to the WAL with CRC32C. Returns nil if no WAL is configured (in-memory mode). Returns ErrWALClosed if the WAL has been closed or is currently being cleared, rather than silently dropping the entry.

func (*WAL) AppendBatch added in v0.6.0 

func (w *WAL) AppendBatch(entries []walEntry) error

AppendBatch writes multiple entries in a single write and single Sync call. Returns nil if no WAL is configured (in-memory mode). Returns ErrWALClosed if the WAL has been closed or is currently being cleared.

func (*WAL) Clear 

func (w *WAL) Clear() error

Clear atomically clears the WAL file. Uses atomic rename to avoid the race between ReadAll (os.ReadFile) and Clear (close+delete+reopen).

func (*WAL) Close 

func (w *WAL) Close() error

Close closes the WAL file.

func (*WAL) ReadAll 

func (w *WAL) ReadAll() ([]walEntry, error)

ReadAll reads all entries from the WAL file (v2 format with magic header and CRC). On CRC mismatch (torn last record), returns the valid records so far and logs a warning. Must hold mu to be atomic with Clear.

func (*WAL) WALPath added in v0.6.0 

func (w *WAL) WALPath() string

WALPath returns the file path of the WAL file. Returns empty string for in-memory stores without a WAL.

Source Files

View all Source files

Directories

Path Synopsis
bench
datasets
Package datasets provides embedded test datasets for MEB benchmarks.
 Package datasets provides embedded test datasets for MEB benchmarks.
cmd
bench command
Command bench runs the MEB benchmark suite and outputs a JSON report.
 Command bench runs the MEB benchmark suite and outputs a JSON report.
Package vector — ADC accumulation dispatch.
 Package vector — ADC accumulation dispatch.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.