v1.2.3 - ABC

[v1.2.3] - ABC - 8/20/2026

v1.2.3 improves Neo4j-compatible Cypher behavior, Bolt transaction safety, retrieval quality tooling, local/remote reranking fixes, and Heimdall provider support.

Highlights

• LiteLLM is now available as a Heimdall chat provider.
• Added reproducible BEIR retrieval benchmarking with recorded SciFact results.
• Restored configurable local and remote reranking.
• Updated llama.cpp to b10411; LLAMA_LOAD_MODE_MMAP is supported.
• Fixed several correlated Cypher aggregation, traversal, list, and subquery behaviors.
• Hardened Bolt explicit transaction lifecycle management.

Retrieval Benchmarks

Recorded SciFact results use the official 300 test qrels with bge-m3:latest at 1,024 dimensions.

Retrieval profile	Recall@100	nDCG@10
BM25 V2	0.88422	0.59974
Accurate HNSW, equal RRF weights	0.93767	0.65534
Exact CPU brute force, equal RRF weights	0.94433	0.67932

Native BGE-M3 reranking was measured over the same 100 candidates:

Profile	Recall@100	nDCG@10	MRR@10	MAP@100
Exact equal RRF, no reranker	0.93563	0.67043	0.64404	0.63486
Exact equal RRF with native BGE-M3	0.93563	0.72292	0.69447	0.69215
Absolute change	0.00000	+0.05248	+0.05042	+0.05729

Reranking preserves Recall@100 because it reorders the same candidates, placing more relevant documents near the top.

These are configuration-specific measurements, not leaderboard claims.

Search and Reranking

• Hybrid retrieval now uses equal RRF weights across query lengths.
• Oversized Ollama embedding input is chunked through the shared chunker before provider evaluation.
• Local GGUF rerankers support rank pooling and classifier output dimensions.
• Query-document pairs use model templates or separator tokens.
• Non-scalar reranker outputs are rejected rather than treated as relevance scores.
• Pooling, attention, context, flash-attention, and provider settings remain configurable for local, Ollama, OpenAI, and HTTP rerankers.

Cypher Compatibility Fixes

• WITH-attached WHERE clauses now correctly evaluate label tests, label conjunctions, boolean combinations, function calls, and null predicates.
• Correlated relationship matches now retain rows through collect(DISTINCT ...).
• List subscripts preserve node bindings through later WITH projections.
• Relationship list properties now work correctly with IN predicates.
• Compound multi-hop patterns now use anchors bound by earlier MATCH clauses.
• Correlated EXISTS and NOT EXISTS now evaluate inner predicates using outer bindings.
• Uncorrelated node-pattern EXISTS subqueries now evaluate correctly.
• Empty Cypher deletes no longer perform unnecessary commit work.

Bolt Transaction Safety

Explicit Bolt transactions now enforce the client-configured lifetime and own a single cleanup attempt across every terminal path.

Key behavior:

• Per-connection executors are supplied by NewWithDatabaseManager or a SessionExecutorFactory.
• A directly supplied TransactionalExecutor remains supported only with MaxConnections: 1.
• Multi-connection servers reject BEGIN when given a shared raw executor.
• BEGIN validates tx_timeout as a PackStream long or null.
• Missing, null, zero, and negative timeout values disable the client deadline, matching Neo4j 5.26 behavior.
• Huge positive timeout values saturate instead of failing.
• Positive lifetimes begin after validated BEGIN reaches backend allocation.
• Transaction expiry cancels active RUN operations and admitted deferred result flushes.
• Expiry triggers one rollback after the operation releases transaction ownership, before responding.
• Cleanup uses an uncancelled five-second request context.
• Cleanup errors, panics, and uncertain commit outcomes close the connection rather than allowing unsafe reuse.
• RESET, ROLLBACK, GOODBYE, and connection loss use the same exactly-once rollback arbitration.
• Persistent Badger/WAL-backed transactions follow the same cleanup rules.
• Deferred PULL and DISCARD flush errors require RESET.
• Pending result and flush state is discarded on terminal paths so only CommitTransaction owns commit durability.

Heimdall

• Heimdall can now use LiteLLM as a chat provider.
• LiteLLM provider initialization and request handling incorporate review feedback.
• The Heimdall watcher once again acquires its pre-execution mutex.

Runtime Updates

• llama.cpp is updated to b10411.
• LLAMA_LOAD_MODE_MMAP configures model loading behavior.

What's Changed

• fix: enforce Bolt explicit transaction lifecycle by @linuxdynasty in #295
• feat: add LiteLLM as a Heimdall chat provider by @prodmanpd in #292
• fix(cypher): evaluate label tests and function calls in a WITH-attached WHERE by @linuxdynasty in #300

New Contributors

• @prodmanpd made their first contribution in #292

Full Changelog: v1.2.2...v1.2.3