Release date: 2026-04-22
Summary
Section titled “Summary”This release adds local and external inference modes, client-side BM25 sparse vectors, bulk insert, recommend queries, script execution, collection dump, and a cleaner error propagation path for hybrid search.
Highlights
Section titled “Highlights”- Local mode —
qql-go connect --inference-mode localgenerates dense and sparse vectors client-side via an OpenAI-compatible embeddings API and upserts explicit vectors to any Qdrant instance. - External mode —
qql-go connect --inference-mode externalfor remote Qdrant + remote embedding endpoints. - BM25 sparse vectors — client-side sparse generation with Rust-compatible tokenization, length-prefixed FNV hashing, and corpus-level BM25 weighting. Stats are persisted in
~/.qql/corpus/<collection>.json. - Auto-probing embedding dimension —
--embedding-dimensionis now optional for local/external mode; the CLI probes the endpoint automatically. RECOMMEND— recommend similar points by positive (and optional negative) example IDs with configurable strategy.RECOMMEND ... OFFSET <n>— pagination for recommendations.RECOMMEND ... SCORE THRESHOLD <f>— filter low-confidence recommendations.RECOMMEND ... WITH { exact: true }— exact KNN baseline for recommendations.RECOMMEND ... LOOKUP FROM <collection> [VECTOR '<name>']— cross-collection recommendations.RECOMMEND ... USING '<vector_name>'— recommend against a specific named vector.INSERT BULK— batch insert multiple documents in a single statement.- Explicit insert IDs —
INSERTandINSERT BULKaccept explicitidfields. USING SPARSE— sparse-only keyword search without dense vectors.CREATE COLLECTION ... USING MODEL— create dense-only collections sized to a specific model.- Script execution —
qql-go execute <script.qql>for running multi-statement scripts. - Collection dump —
qql-go dump <collection> <output.qql>for exporting collections as runnable QQL scripts. - REPL shortcuts —
\e <file>and\dump <collection> <file>for script execution and dump from the interactive shell. - Connect-time endpoint validation — the embedding endpoint is test-called during
connectto surface misconfiguration early.
Known limits
Section titled “Known limits”Full changelog
Section titled “Full changelog”See CHANGELOG.md.
For full commit list and details see the original release notes.