AntflyAntfly

Changelog

All notable changes to Antfly will be documented in this file.

Releases#

[Unreleased]#

  • Trusted-principal authorization — reject requests carrying a trusted principal when no verifier is configured, including MCP and inference routes with local authentication disabled. MCP sessions cannot preserve a previous request's broader permissions.

  • PostgreSQL integration naming — rename the Cargo package from pgaf to antfly-postgres, the PostgreSQL extension and query-builder schema to antfly_postgres, and move its source to rs/crates/postgres. Enable it with CREATE EXTENSION antfly_postgres;.

  • Native PDF extraction — repair demonstrably interleaved horizontal prose columns without changing paint order. Require sustained prose in both columns and preserve ambiguous table row associations. Preserve authored order for overlapping paint sequences, already-grouped columns, and vertically separated layouts. Resolve indirect CCITT decode parameters, accept valid zero-filled Group 3 line markers, and correct PNG compression-buffer ownership.

  • Durable OCR grounding — preserve compatible Reader regions as UTF-8 byte spans and page-space boxes, including rotated pages and CropBox offsets. Invalid regions no longer discard valid text or shift later repeated text onto an earlier occurrence.

  • OCR content retention — reject partial transcriptions that discard most decoded text while allowing recovery from corrupted glyphs. Preserve signed numeric values, singleton numeric lines, and their full source-row context when combining OCR with embedded text.

  • Native multistage OCR — support exact transposed convolution and average pooling, preserve typed constants and dynamic reshape dimensions in native partitioned execution, and correct Paddle image normalization, detector aspect ratio, recognition width/padding, tall-crop rotation, and CTC vocabulary handling. Recognition failures propagate instead of returning partial success.

  • Reader request memory — encode large inline-image JSON requests into one exactly sized allocation, avoiding unnecessary working-set failures without increasing production memory limits.

  • Bounded compatibility rendering — keep temporary RGBA and PNG compression allocations out of retained-output budgets without increasing memory limits. macOS compatibility-rendered pages remain explicitly marked as such.

  • Research agent — POST /agents/research plans a question into sub-questions, researches them in parallel with the retrieval agent, reflects on coverage, and writes a sectioned report whose [E#] citations resolve to a deduplicated evidence registry. Unresolvable citation markers are removed and reported, and an optional verify step checks cited statements against their evidence. Every limit is declared in budget and is cumulative across resumes; research_state carries a run forward. Durable jobs (/agents/research/jobs, with advance and cancel) checkpoint after every phase, are scoped to the authenticated user, and survive restarts in local standalone mode. Available from the CLI (antfly agents research, --job), the Go and TypeScript SDKs, @antfly/components (useResearchStream, ResearchReport), evalaf, and as the A2A research skill.

  • Retrieval agent fetch tool — agents can read full web pages. Fetch is opt-in and admits only URLs returned by web_search in the same run or on fetch_config.allowed_hosts; private addresses are always blocked. Tool results are budgeted in estimated tokens instead of bytes, and every model round checks the request deadline and cancellation.

  • One rerank endpoint for text and images — POST /ai/v1/rerank takes documents, where each entry is a string or an array of text and image content parts, the same format generation and embedding use. Documents with images need a multimodal reranker (ColQwen or Qwen3-VL); other models reject them with 400. /ai/v1/rerank_multimodal is removed; send those requests to /rerank with the document content directly in documents (the {id, content} wrapper is gone). prompts is deprecated but still accepted for text. The operation id is now rerankDocuments, so generated SDK methods are renamed (RerankDocumentsWithResponse, rerank_documents). The Go SDK adds InferenceClient.RerankMultimodal, and the TypeScript rerank() accepts content-part documents. Inference proxy route rules that matched the rerank_multimodal operation should match rerank and select the model by name instead. Inference servers advertise rerank_documents_v1 in the model catalog; Antfly sends documents only to servers that do and keeps sending prompts to older ones, so clusters and inference pools can upgrade in either order. prompts can be removed once every client sends documents.

  • Embedder input types come from model capabilities — the embedder multimodal flag is removed. Antfly learns which inputs a model accepts from its capabilities (discovered from Antfly inference, or the linked runtime). For a model whose capabilities cannot be discovered yet, set inputs, for example ["text", "image"], which replaces the discovered input types; only the antfly (image, audio) and bedrock (image) providers accept media inputs. Existing indexes keep working: a stored multimodal value is ignored, and it no longer takes part in embedding producer identity.

  • Reranker image support is resolved from the model manifest — like image embedding, a manifest declaring text-only inputs stays text-only, and otherwise a Qwen3-VL GGUF bundle with its projector or a colqwen / multimodal_late_interaction capability selects the image executor. The model catalog, the executor contract, and the rerank handler share that one answer.

  • Query rerankers score images — a reranker template can use the media and remoteMedia helpers, and each candidate's images are sent with its text to an Antfly reranker whose model accepts images, whether it runs in process or on a remote inference server. Other rerankers reject such queries with 400.

  • Aggregations on hybrid and semantic queries — queries that combine aggregations with semantic_search or embeddings no longer fail with query_candidate_budget_exceeded. Aggregations now count every full-text match plus each embedding index's top results at the requested limit and search_effort, merged across shards. Previously the internal collection widened each vector search to the full-result budget, which approximate search could never prove complete. The ranked page is unchanged, the full-result budget still bounds full-text matches, and a filter with no full-text query contributes no extra rows. Separately, a vector search that exhausts its traversal with a truncated top-k window now reports its total as a lower bound instead of exact.

  • Aggregation and semantic paging corrections — full-text facets with fewer matches than the collection budget now complete on larger tables. Explicit match_all contributes its text domain to hybrid aggregations, while filter-only vector requests no longer add a match-all retrieval lane to the ranked page.

  • TOON document rendering in retrieval agents — the generation prompt now encodes each retrieved document's fields as TOON instead of raw JSON, and document_renderer on a retrieval agent request is accepted again: a Handlebars template rendered per hit against {id, score, fields} with an encodeToon helper (indent, delimiter). Invalid templates are rejected with 400. Queries still reject document_renderer.

  • antfly inference run reads the rest of its config file — keep_alive (duration) or keep_alive_ms (integer; 0 never unloads idle models), prompt_cache, and kernel_jit are now honored, nested under inference or flat. CLI flags and ANTFLY_INFERENCE_KERNEL_JIT_MODE still win. Operator InferencePool keep-alive settings now take effect.

  • Operator: standalone keeps embedded inference — standalone clusters no longer default inference.api_url to http://0.0.0.0:11433, which disabled embedded inference; api_url is written only when set explicitly, and a user-provided value is never overwritten. Standalone and InferencePool pods with a memory limit get ANTFLY_PROCESS_MEMORY_BUDGET_MB at 90% of it.

  • Operator: InferencePool model priority orders preloads — eager preloads warm high, then medium, then low priority models; priority does not affect eviction.

  • Retrieval agent auto_seed stays agent-side — the flag is cleared before the internal query hop now that the generated rerank type carries it.

  • antfly standby replaces antfly ha — the hot-standby command is renamed; antfly ha remains a hidden alias for one minor release. It gains --data-dir (opens the node's standby state under <dir>/ha/ and reads the log identity from the files), --config (reads the server's hot_standby section), unprefixed flags (--admin-url, --admin-token-env, --cluster-id, ...; the --ha-* spellings remain aliases), and ANTFLY_STANDBY_ADMIN_URL / ANTFLY_STANDBY_ADMIN_TOKEN defaults (the ANTFLY_HA_* names are read as fallbacks). The -- separator before the verb is optional. Help text rewritten.

  • hot_standby: config section — antfly standalone --config fills any --ha-* flag it was not given from hot_standby.{admin,identity,primary,standby,sync,retention}; flags still win. The deprecated ha: key is still read for one minor release.

  • Admin API moved to /admin/v1/standby/... and /standby/v1/..., and the internal replication API to /internal/v1/standby/replication/.... The standby-role routes drop their now-redundant segment (/standby/status, /standby/bootstrap, /standby/upstream). The old /admin/v1/ha/..., /ha/v1/..., and /internal/v1/ha/... paths are served as aliases for one minor release, and the CLI and a replicating standby fall back to them once when they meet a 0.2 server, so either side of a pair may be upgraded first. The Kubernetes operator negotiates per server (--standby-admin-path-style, default auto) instead of waiting for a minimum server version, and reads its own token from ANTFLY_STANDBY_ADMIN_TOKEN before ANTFLY_HA_ADMIN_TOKEN.

  • Schemas renamed HA* -> Standby* — in the admin and internal OpenAPI specs and the generated SDKs. The Go SDK keeps every HA* name as a deprecated alias, adds PathStyleAuto (canonical first, one fallback to the old spelling, remembered per base URL), and keeps the legacy style as its zero-value default for this release.

  • Operator metrics — dual-emitted under both the ha and standby subsystems during the deprecation window.

  • antfly standalone --hot-standby-* flags replace --ha-*: --ha-standby-X becomes --hot-standby-X (--hot-standby-log, --hot-standby-progress, ...), --ha-primary-X becomes --hot-standby-primary-X, and every other --ha-X becomes --hot-standby-X. The --ha-* spellings remain aliases for one minor release; the operator keeps generating them for now.

  • Data directory standby/ — antfly standby --data-dir now looks for standby/{primary.wal,slots,log.wal,progress.wal,fence.wal} and falls back to the pre-0.3 ha/ tree. The server migrates an ha/ tree to standby/ once at startup when its flags point at the new tree (directory rename plus standby.wal -> log.wal, standby-progress.wal -> progress.wal; nothing is deleted or overwritten). The operator switches a cluster's default pod paths to /antflydb/standby/ once it has seen the cluster's nodes speak the 0.3 admin API and records the choice in status.haStatus.dataLayout; new clusters start on standby/.

  • Server metrics antfly_standby_* — every antfly_ha_* series is also emitted under the new name for one minor release.

  • OpenAPI tags and config schemas — tags standby / standby-replication; HotStandbyPrimaryConfig / HotStandbyStandbyConfig.

  • storage/hot_standby package — the Zig package moved from storage/ha, the command source from cmd/ha.zig to cmd/standby.zig, and the root alias from antfly.ha to antfly.hot_standby; test names and build steps are storage.hot_standby ..., standby cmd ..., and antfly-storage-hot-standby-*.

  • Planned switchover — antfly standby switchover --to <standby> fences the old primary first, waits for the standby to reach the final LSN, fences and promotes the standby with the same fence generation, assesses the old primary's rejoin (reseed applied immediately, rewind pending its restart as a standby), and repoints --follower standbys.

  • antfly standby follow and POST /admin/v1/standby/upstream — repoint a running standby at a new primary without a restart, guarded by an expected identity precondition; idempotent on retry. Also fixes a race where the replication round read its upstream config outside the standby state mutex.

  • Fence generation allocation — POST /admin/v1/standby/fence no longer requires generation; without a Kubernetes Lease authority the node allocates the next generation, and an identical omitted-generation retry returns the held receipt instead of double-fencing.

  • The ha cmd test root now runs all of its tests (the filter previously ran 7 of 35); tests that wrote to stdout under the build runner were the cause of a hung zig build.

  • Relational packed base-row storage lands behind the relational table profile

  • HBC vector LSM replaced with native WAL-backed generations

  • Self-contained deterministic VOPR testing, with production HA promotion and scaling composed in simulation against retained corpora

  • Bounded multimodal document processing in inference

  • Homebrew version resolution and macOS dylib relocation made reliable

  • Docs for the v0.2 launch rebuilt on one scaffold with facts re-verified against main

[0.2.1] - 2026-09-08#

Not just a patch release. v0.2.1 fills in parts of the v0.2.0 release that shipped rough or missing, on top of a long list of bug fixes found in the first weeks of the Zig runtime in production.

Highlights#

  • Hot-standby HA — standalone mode now has a Postgres-style hot standby with fenced promotion, a lower-cost alternative to Raft replication. Supported end to end by the operator and CLI, and chaos tested for weeks @bpopadiuk
  • Generated GPU kernels — a Triton-style quant-kernel JIT now generates our CUDA and Metal kernels, accelerating Gemma 4, BGE-M3, and GLiNER2 while cutting the number of hand-maintained quantization types @timkaye11
  • Qwen3 across the board — Metal and CUDA support for Qwen3-Embedding and Qwen3-VL; Qwen3 is now the suggested embedder, reranker, and multimodal reranker, with a q8_0 bundle in the quickstart sized for a laptop @timkaye11
  • PDFs, properly — native rendering, OCR handoff, whitespace reconstruction, JBIG2 and JPEG 2000 decoding, and ParseBench-guided parser fixes, validated against benchmark corpora @dovinmu
  • Benchmarked — graph, full-text, and vector index behavior benchmarked extensively, and many of the metrics you care about improved as a result. A release "model card" with our evaluation suite is coming @dovinmu
  • Structured graph query DSL — a Cypher-like DSL covering a large subset of Cypher functionality, with much faster exact pattern queries
  • Multi-source vector indexes — index multiple fields with the same embedding type into a single index

Features#

  • Structured graph query DSL, plus unit hierarchy grouping and traversal in queries
  • Multi-source artifact indexes restored to production grade
  • Managed embeddings published progressively, so partial indexes stay queryable during enrichment retries
  • Authoritative cross-runtime index readiness, recoverable index generation repair, and a clearer index wait
  • Catalog routing as a first-class capability
  • Algebraic indexes gain dynamic projections and adaptive HyperLogLog cardinality
  • $min document transforms
  • Object storage checksum provenance; embedding activity and backup bundle architecture exposed
  • Lake-native Parquet and Iceberg serving foundations, and a yacc-generated SQL parser foundation
  • Configurable query, write, and inference admission; adaptive vector cache governance; unified inference resource ownership with enforced process envelopes
  • Gemma 4 26B-A4B on Metal and CUDA; faster Gemma 4 E2B/E4B Metal decode with long-context correctness fixes @timkaye11
  • Metal support for ModernBERT, NomicBERT, BGE-M3, Florence-2, and GLiNER2 and DeBERTa rerankers; quantized Mixedbread ONNX rerankers @timkaye11
  • Inference execution is cancellable and observable, with HTTP cancellation and worker recovery tested end to end
  • Operator: custom Kubernetes cluster domains, stable data pod routes, and unsafe metadata replica changes rejected
  • Standard Linux containers built with glibc; CLI packages published from hosted runners; Python 3.11 through 3.14 supported

Bug Fixes#

  • Raft: retry-safe data apply replay, progress rounds and exact metadata proposal receipts, isolated read-index heartbeat contexts, bounded replication pipelines, and preserved metadata cadence
  • Restore: fenced restore attempts recover, CLI restore polling tolerates transient responses, generated indexes survive native backups, and backup create admission is enforced
  • Metadata: exact proposal application, mutation leader discovery, control-round authority churn, and fenced snapshot reads
  • Index lifecycle: false DDL failures after status fencing, deferred lifecycle retries, stale incarnation status handoff, and repair handoff through fenced status publication
  • Enrichment retry wedges across runtime boundaries; artifact-backed embedding indexes executable; artifact indexes preserved during embedding publication; materialized chunk embedding writes after list growth @dovinmu
  • Artifact full-text merge scaling past 1M chunks
  • Auth: request authorization boundaries hardened, durable destination authority bound, internal RPC authenticated across upgrades, transaction leaders rediscovered securely, direct browser query access secured
  • HTTP: 413 returned before admitting oversized HTTP/1 bodies
  • Inference: Bedrock request construction, preload registry variants, POSIX file advice gated to Linux, weight lifetimes bounded and Metal request contexts reclaimed, embedding CLI results written to stdout, standalone workers isolated
  • PDF: compressed remote content decoded before extraction @dovinmu, stencil colors and sparse pattern transparency honored, remaining pdf-hard corpus gaps closed
  • Autosplit: lost forced scans, and incomplete split provisioning snapshots retried
  • Quickstart: index readiness, search, chunking, and retry contracts; Gemma multimodal inference and streaming completion
  • Antfarm: table creation defaults and errors, artifact query retrieval
  • Chunk overlap bounded and generation capacity errors preserved
  • HBC recall under cache governance; overlapping vector cache fills bounded

Performance#

  • Exact graph pattern queries optimized
  • Persistent ordered memory snapshots; public vector ingest memory and compaction optimized; HBC vector cache reads sharded
  • Linear merge uses ordered hash scans; catalog routing no longer takes the Raft runtime lock
  • Bounded runtime graph and parallel memory scheduling ported to the Zig runtime

Removed#

  • The end-of-life Go server implementations. The Go SDK, operator, docsaf, evalaf, genkit plugin, and memoryaf remain

Packaging#

  • Go, Python, and TypeScript SDKs v0.2.0 released to coincide with the v0.2.x series
  • Operator v0.1.0 supports the full v0.2.x feature set and CLI interface
  • Antfly installs with pip, npm, or Homebrew, and inference on macOS is much friendlier

Thanks to @markhayden, James, and Drew for finding and helping tackle things that were dropped from the Go port or were plain bugs in the new code, and to batz.prime for finding restore bugs on Raft deployments and the testing gaps behind them.

Full changelog

[0.2.0] - 2026-08-10#

Antfly v0.2.0 is a ground-up rewrite of the engine in Zig with zero dependencies outside the repository. It is the default release from this version on, uses less memory and CPU than the Go engine with a smaller RSS footprint, and adds a great deal of surface area that the Go engine never had.

Highlights#

  • Zig runtime — lower memory and CPU usage, a smaller RSS footprint, zero third-party dependencies, and WASM compatibility
  • Storage backends for every workload — single-file (Antfly Lite, .aflite, SQLite-shaped), standalone LSM, distributed multi-Raft, and serverless over object storage, with portable AFB backups to move data between them
  • Inference in the database — run GGUF, ONNX, and safetensors models, or fine-tune them, directly in Antfly, across many LLM and small-model architectures with Metal and CUDA acceleration
  • Artifacts API — declarative JSON pipelines that run OCR, graph relation extraction, embeddings, LLM enrichment, and classification inside the database, the work you would otherwise build as a downstream queue or data engineering pipeline
  • Query language — a simpler Bleve-style DSL that covers roughly 99% of Elasticsearch query capabilities, plus an improved foreign table planner
  • Index management — Postgres-style visibility and control over indexes: rebuilds, cursors, and health status
  • Wider content types — native PDF to PNG rendering for embedding generation, more audio formats, and hardened remote content handling
  • Testing — deterministic simulation (VOPR-style) and TLA+ trace validation across Raft, transactions, snapshots, and shard splits

Upgrade compatibility#

Antfly 0.2.0 is not an in-place data-directory upgrade from 0.1.x. Durable formats changed without a 0.1.x compatibility path. Replace or wipe 0.1.x pods and create fresh 0.2.0 data directories; export and restore any data that must be retained.

The Python SDK now exposes transform operators with stable semantic enum names (SET, SET_ON_INSERT, UNSET, INC, ADD_TO_SET, and MAX). Positional VALUE_n names are not part of the 0.2 API.

Dense-query _score values now follow the same relevance convention as text, sparse, hybrid, and Elasticsearch-style search results: higher values rank first. Code that previously treated dense _score as a lower-is-better vector distance must use the new optional _distance field instead. _distance retains the raw metric-specific distance for direct dense hits. On a source group ranked by dense descendants it is the distance of the best descendant that supplied the group score. It is omitted for non-dense and fused results.

Features#

  • Antfly Lite: the engine as a single embedded file
  • Hot-standby HA design and runtime surface (productionized in 0.2.1)
  • Extension system v1, with extension MCP tools filtered by permission
  • DOCID document identity and query APIs; derived document hierarchy in Antfly and docsaf
  • Artifacts API, artifact CLI, and enrichment listing; enrichment assets routed into full text
  • GET /connections listing connected providers, models, and stores; cluster topology status endpoint
  • Data-plane auth CLI commands; trusted-principal MCP auth; authentication required on the public MCP endpoint; A2A and internal retrieval routes admin-only @dovinmu
  • Retrieval agent tools configurable per request step; fielded and raw MCP query requests
  • Native Gemini and Bedrock multimodal embeddings, including Titan @excenter; Jina v5 direct Metal embeddings; unified managed embedders
  • GLiNER2: relation extraction, ONNX exports and symbolic Metal fallbacks, split GGUF bundles, and resident fine-tuning on Metal; LoRA policy alignment across fine-tuning recipes @timkaye11
  • Gemma 4 E2B, E4B, and 12B on Metal and CUDA with production tool calling, MTP, QAT, and unified image and audio inference @timkaye11
  • Florence-2 on Metal and CUDA; Tensor Core CUDA dispatch for CLIP/CLAP, mxbai rerankers, and GLiNER2 @timkaye11
  • Native traditional ML predictors with antfly inference pull and /ml/v1/predict @timkaye11
  • Inference CLI consolidated under antfly inference; SDK model pull helpers; inference defaults
  • Antfarm: design system rebuild @stinkbugaf, Graph Index Explorer @timkaye11, connections view, simplified navigation, hosted mode
  • Operator: Antfly and Termite operators consolidated, autoscaling with storage scale-down safety, single-node swarm mode @timkaye11, serverless operator and proxy, reduced RBAC @dovinmu
  • docsaf: Google Drive auth, derived document hierarchy
  • Entity-resolution design and comparison-levels scorer; tenant-scoped ARD discovery surfaces
  • pip and npm CLI packaging; Antfly Cloud CLI shim @dovinmu; index query latency metrics
  • Zig runtime, proxy, operator, Termite, and SDKs consolidated into the monorepo; Go fallback artifacts no longer built during release

Bug Fixes#

  • Composed vector query correctness, with faster hybrid filtering
  • Raft: apply deadlock behind nightly e2e hangs, HTTP snapshot catch-up, data leader readiness publication, metadata bootstrap churn and no-op persistence, startup catch-up busy loop
  • LSM: root and WAL locking hardened, recovered temp files cleaned before WAL replay, bloom filters stored only in SSTables
  • Backup: restore writer conflict, backup forwarding, backup job URL configuration, read-index barrier for metadata backups, Go portable AFB backups restorable in Zig @dovinmu
  • Dense replay catch-up convergence and progress
  • Writes outside the byte range rejected during split transition; large linear merge requests handled
  • TTL metadata for local document writes; query aggregation label lifetime; bool must+filter clause parsing @dovinmu
  • Zig OOM retirement lifecycle cleanup @dovinmu; partial config defaults
  • ONNX and Metal: dependency-free ONNX loading, symbolic projection embeddings and runtime shapes, quantized GGUF embedding lookup shapes @timkaye11, CLIP preprocessing contract and Metal parity, clipclap ONNX correctness, Metal SDPA layout, macOS Metal builds without MPSGraph @timkaye11
  • PDF: malformed content recovery, extraction failure recovery, bounded extraction memory, unsupported OCR backend paths @stinkbugaf
  • Corrupt indexes dropped without blocking shard startup; full-text enrichment index provisioning; artifact-backed chunk embeddings
  • Remote content credential resolution and security config; store backup path handling @dovinmu; SearchAF Zig runtime response contracts
  • TypeScript SDK: openapi-fetch bundled to fix a CJS interop crash @markhayden; vulnerable TypeScript and Go dependencies bumped @dovinmu

Performance#

  • Read-path scalability, quarantine self-heal, and HBC link repair
  • Live-doc visibility, search effort policy, and stale payload repair @dovinmu
  • LSM online write and bulk ingest tuning; load memory pressure reduced; metadata Raft memory retention bounded
  • Gemma 4 E2B Metal reuse and CUDA inference speedups @timkaye11

Removed#

  • MLX backend

Full changelog

[0.1.1] - 2026-03-31#

Highlights#

  • Faster Standalone Mode — local shard bypass eliminates network hops and JSON serialization for single-node deployments, significantly reducing query latency
  • Smarter Vector Search — automatic reranking and size-aware search effort tuning deliver better recall out of the box
  • Friendlier Errors — LLM generation failures now return clear, actionable error messages with appropriate HTTP status codes
  • Automatic Shard Management — Antfly now automatically splits large shards and merges underutilized ones, keeping cluster performance balanced without manual intervention
  • memoryaf — a new MCP-compatible memory service with built-in HTTP API and dashboard for memory-augmented search

Features#

  • Automatic shard split policy keeps individual shards from growing too large
  • Online shard merges consolidate underutilized shards without downtime
  • Local shard bypass skips HTTP and serialization overhead in standalone mode
  • Single-shard fast path and index caching speed up common query patterns
  • Automatic reranking for HBC vector search improves recall without configuration
  • Size-aware search effort defaults adapt to index size automatically
  • Hybrid full-text fallback when embedding indexes are unavailable
  • Packed dense and sparse embedding format accepted in the query API
  • Unified bounded agent APIs across all generated clients
  • memoryaf: MCP-compatible memory service with HTTP API and embedded dashboard
  • Richer query builder schema context for better autocomplete
  • NER entity extraction and knowledge graph support in docsaf
  • LLM generation error classification with user-friendly messages @dovinmu

Bug Fixes#

  • Fix cosine distance centroid normalization in HBC indexes @dovinmu
  • Fix RaBit centroid clone aliasing causing quantization regressions @dovinmu
  • Fix nil-interface panic in enricher stats during health checks — thanks @montanaflynn!
  • Fix raft serveChannels shutdown hang — thanks @montanaflynn!
  • Fix shard orphaning in StopRaftGroup when shard is initializing @esniff
  • Fix Pebble close races during split finalize and node shutdown
  • Fix split write routing during handoff, cutover, and readiness transitions
  • Fix Store.Close goroutine leak after shard shutdown
  • Fix index cache concurrency with proper LoadOrStore semantics
  • Fix SSE error events not being sent when RAG generation fails @dovinmu

Performance#

  • Eliminate JSON serialization on the local shard search path
  • Optimize inter-node vector search serialization

Full changelog

[0.1.0] - 2026-03-17#

Highlights#

  • First official public release of Antfly

Full changelog

[0.0.19] - 2026-03-17#

Highlights#

  • Pruning Fix — handle negative scores in pruning
  • Build Version — add build version ldflags

Full changelog

[0.0.18] - 2026-03-16#

Highlights#

  • Initial Public Release — Antfly is now open source
  • Transaction Safety — multiple fixes for nil pointer panics during shutdown and transaction recovery
  • Schema Migration — table migration state exposed via API for schema version cutover
  • SQL Join Fix — LEFT JOIN no longer returns INNER JOIN results due to plan cache key collision

Features#

  • Initial public release of Antfly
  • Add table migration state to API for schema version cutover
  • Replace map[string]any transaction records with typed TxnRecord struct
  • Thread commit_version through ResolveIntentsOp proto and clean up transaction DB code
  • Fix HBC delete to repair underfull leaves and update recall expectations
  • Fix race in Group causing ResultGroup to drop results from queued tasks
  • Fix nil pointer panic in transaction recovery during shutdown
  • Guard transaction Pebble accesses with pdbMu to prevent nil dereference on shutdown
  • Fix LEFT JOIN returning INNER JOIN results due to plan cache key collision

[0.0.17] - 2026-03-15#

Features#

  • CLIPCLAP multimodal embedder capabilities (#421)
  • search_after/search_before cursor pagination (#413)
  • Enrichment pipeline error handling for broken remote resources

[0.0.15] - 2026-03-10#

Highlights#

  • Foreign Table Joins — query across foreign tables with automatic filter pushdown and SQL aggregations
  • Routed PG Replication — route PostgreSQL CDC streams to specific tables with a new evaluator package for custom replication logic
  • Distance Metric Configuration — choose between cosine, euclidean, and dot-product distance per embedding index
  • Index & Enricher Observability — new stats endpoints exposing document counts, enrichment progress, and per-index health

Features#

  • pgaf (now antfly-postgres; SQL extension antfly_postgres) — PostgreSQL extension providing a custom index access method (CREATE INDEX ... USING antfly), @@@ operator for full-text/semantic/hybrid search, query builder functions, sync triggers, and antfly_search() for native Postgres integration
  • Foreign table join support with filter pushdown and SQL aggregations (#400)
  • Routed PostgreSQL replication with evaluator package extraction (#403)
  • Configurable distance metrics for embedding indexes (#406)
  • Comprehensive index and enricher stats API (#405)
  • CLIP image search improvements for multi-image queries (#406)

[0.0.13] - 2026-03-03#

Highlights#

  • Sparse Vector Search (SPLADE) — hybrid search combining dense and sparse vectors with weighted fusion for better relevance
  • PostgreSQL CDC Replication — automatically sync data from PostgreSQL into Antfly via logical replication
  • Operator Improvements — PVC lifecycle management, availability zone topology, admission webhooks, and storage resilience
  • Faster Sparse Indexing — up to 3.6x faster sparse index inserts with multi-tier caching

Features#

  • Sparse Vector (SPLADE) Search — hybrid dense+sparse fusion with configurable per-index merge weights
  • PostgreSQL CDC Replication — logical replication with automatic change capture from PostgreSQL tables
  • Chunking and Summarization support in sparse embeddings index
  • Operator: PVC Lifecycle management, availability zone topology, and storage resilience
  • Operator: Admission Webhooks for AntflyCluster resource validation

Performance#

  • Up to 3.6x faster sparse index inserts
  • Multi-tier caching and batched writes for sparse indexes
  • Configurable sync level for embeddings indexes

[0.0.9] - 2026-02-22#

Highlights#

  • Secrets Management — new API and dashboard page for managing secrets
  • API Key & Bearer Token Auth — authenticate with API keys or bearer tokens
  • AI Provider Timeouts — configurable timeout for AI provider calls
  • PDF Enrichment Pipeline — zip-direct reading, parallel extraction, vision-based categorization, and Florence 2 re-OCR for low-quality pages
  • Omni Edition — renamed install edition with streamlined macOS support

Features#

  • Secrets Management API and Antfarm dashboard page
  • API Key & Bearer Token Authentication
  • AI Provider Timeout configuration
  • PDF Enrichment — direct zip reading, parallel page extraction, page-type categorization with vision support, and Florence 2 re-OCR for low-quality pages
  • Dashboard improvements with reverse proxy support and sidebar redesign

[0.0.8] - 2026-02-18#

Highlights#

  • Cross-Table Transactions — optimistic concurrency control (OCC) with read-modify-write support across tables
  • Built-in Embedder & Reranker — bundled INT8 quantized all-MiniLM-L6-v2 embedder and reranker, no external service required
  • Shared Pebble Block Cache — single block cache shared across all DB instances per process for better memory utilization
  • Operator Scheduling Constraints — tolerations, nodeSelector, affinity, and topologySpreadConstraints in AntflyCluster CRD
  • Cluster Hibernation — scale operator replicas to zero while retaining PVCs for cost savings

Features#

  • Cross-Table Transactions with OCC read-modify-write support
  • Built-in Embedder — INT8 quantized all-MiniLM-L6-v2 bundled with Antfly
  • Built-in Reranker — INT8 quantized reranker model bundled with Antfly
  • Shared Pebble Block Cache across all DB instances per process
  • Operator Scheduling Constraints — tolerations, nodeSelector, affinity, and topologySpreadConstraints added to AntflyCluster CRD
  • Cluster Hibernation — allow scaling metadata and data node replicas to zero

[0.0.7] - 2026-02-13#

Highlights#

  • Retrieval & Generation Agents — new agentic architecture for retrieval-augmented generation
  • MCP & A2A Protocol Support — connect Antfly to AI agents via MCP (/mcp/v1) and Agent-to-Agent protocol
  • Foreign Tables — federated queries against external PostgreSQL databases
  • Named Provider Registry — configure embedders, generators, rerankers, and chunkers by name
  • Audio Transcription — speech-to-text support via Termite
  • Ephemeral Chunks — transient chunk storage with the store_chunks config option

Features#

  • Retrieval Agents — tool-use agentic loop for retrieval and generation, replacing the previous answer endpoint (deprecated /agents/answer still available for backward compatibility)
  • MCP Server at /mcp/v1 for AI agent integration
  • A2A Protocol facade for retrieval and query-builder agents
  • Foreign Tables for federated PostgreSQL queries
  • Named Provider Registry for embedders, generators, chains, rerankers, and chunkers
  • Audio/STT with Termite as speech-to-text provider and media chunking support
  • Ephemeral Chunks mode (store_chunks config option) for transient chunk storage
  • Graph Index — field-based edges, topology constraints, and summarizer
  • Remote Content configuration system for web scraping
  • CLAP & CLIPCLAP model support for audio embeddings
  • Antfly Operator now included in the main repository with docs and install manifests

[0.0.2] - 2026-01-10#

Highlights#

  • Cross-table join support for queries spanning multiple tables
  • Zero-downtime shard splitting with two-phase split
  • TTS/STT audio library with OpenAI and Google Cloud providers
  • CLIP model support for multimodal image indexing

Features#

  • Cross-Table Joins — query across multiple tables with shard-aware routing
  • Zero-Downtime Shard Splitting — two-phase split for high availability
  • Audio Library — TTS and STT support with OpenAI and Google Cloud providers
  • Dynamic Templates — flexible field mapping with automatic schema inference
  • Aggregations API — range and term aggregations (renamed from Facets)
  • Chat Agent — tool execution, clarification handling, confidence scoring, and multi-turn query builder mode
  • Indexes in Raft Snapshots — faster recovery with pause/resume for index operations
  • ONNX Runtime GenAI bundled for local LLM generation

[0.0.1] - 2025-12-20#

First official release of Antfly.

Features#

  • Unified ONNX + XLA build for cross-platform ML inference
  • Termite downloads page with Homebrew support
  • antflycli included in container images
  • Document TTL support