Changelog

What's New in Optiq

Every update, improvement, and fix we ship. Follow along as we build.

April 17, 2026

Security

·Per-repo AES-256-GCM encryption keys generated on your device — sent to server during indexing for encryption, then discarded
·API responses always return encrypted code content — decryption happens exclusively on the CLI
·Encryption keys stored locally at ~/.optiq/keys/ with restricted file permissions
·Persistent audit log tracking logins, role changes, key creation, and account deletions
·GDPR-compliant account deletion — all user data purged across every table

Performance

·Indexing 7x faster — embedding batch size increased, file nodes skipped, backfill eliminated
·Thread-safe DuckDB with connection-per-thread pooling
·Database indexes on all hot lookup columns for faster auth and queries
·LRU engine cache with configurable max size (prevents OOM on large deployments)
·Structured JSON logging with request ID tracking across all log entries
·systemd service management with auto-restart and boot-on-startup
·Automated hourly database backups with 7-day retention

Platform

Engine

·Full C# language support — parsing, chunking, symbol resolution, import tracking
·optiq_search now accepts a path parameter for direct directory-based search
·Repo ID sanitization for folder names with spaces and special characters
·Status page now shows service data to all visitors

April 6, 2026

Performance

·Search engine hot path rewritten in Rust via a new optiq-core native module
·Engine startup reduced from 36s to ~3.5s for a 20K-node repo
·Cold search dropped from 46.7s to ~2.4s, warm searches complete in ~1.7s
·Advanced ranker initializes 5x faster using parallel tokenization via rayon
·Bulk decryption is 16x faster
·Binary graph cache enables near-instant subsequent server restarts instead of rebuilding from the database every time

Retrieval

·Flow-intent queries now surface related code across files — searching "how are slash commands registered and executed" correctly returns both registration-side code and execution-side code
·Query classifier detects flow-intent queries and enables graph expansion, following call edges between files to find related code that pure semantic search misses
·Reranker filters out false positives like unrelated files that happen to share a filename keyword

April 3, 2026

Retrieval

·Search queries now understand domain synonyms — searching "how does authentication work" finds functions like verifyToken, checkAuth, and loginUser even if they never mention the word "authentication"
·The engine detects what kind of query you're running and adjusts its strategy. Identifier searches lean on exact keyword matching, abstract questions lean on semantic understanding
·Queries asking "where is X" now return the actual file alongside individual functions instead of dropping file-level results
·Reranking prioritizes functions that directly implement the queried feature over generic utilities, formatters, loggers, and trivial wrappers

Quality

·Auto-generated code chunks now receive descriptive, context-aware names instead of generic labels like block_1 or handler_2
·Fixed duplicate node extraction — inner functions inside large components were being indexed as standalone symbols, causing the parent's logic to get dropped during dedup
·Fixed double-extraction of exported functions that were being indexed twice
·Named sub-chunks no longer incorrectly removed as "redundant" — only auto-generated block/chunk labels are filtered
·Reranker now sees the actual matched code region instead of the first few boilerplate lines of large promoted functions
·Fixed HNSW vector index key mapping that could return the wrong node

March 31, 2026

Retrieval

·Fixed a critical bug where keyword search was completely dead for all repos — every query was running on semantic similarity alone with zero keyword/BM25 signal
·Root cause: the full-text index was never built during indexing, only on a close() call that never happened
·All existing repos have been rebuilt with working keyword indexes
·Query terms now match against filenames — searching "pricing" boosts Pricing.tsx even in natural language queries
·Stale or schema-mismatched indexes are now auto-detected and rebuilt on first search instead of silently failing

MCP

·optiq_search now automatically detects the current repository instead of requiring the agent to specify it on every query

March 30, 2026

Dashboard

MCP