v0.3.0 — Two-stage Jev retrieval and task model configuration

Two-stage Jev retrieval

Jev now selects graph relations and reranks their source passages together with direct vector-search candidates. Enable it with reranker_model="jev"; no additional installation extra is required. Explicit Jev versions and the jev-latest / jev-preview aliases are supported.

Across 1,000 MuSiQue and 1,000 2Wiki queries, the two-stage pipeline reaches 86.07% average Recall@5 and the quality–latency Pareto frontier among the compared methods under the reference latency estimates. See the evaluation report.

Model configuration

Configure extractor_model, reranker_model, and answer_model independently, with llm_model retained as the shared default. ModelConfig supports task-specific endpoints and credentials. Existing string model configuration and the default generative reranker remain supported.

Fixes and compatibility

• Preserve selected relation order when expanding to passages. This corrects ranking behavior and can change retrieved passage order.
• Keep /settings compatible with the frontend when llm_model uses ModelConfig.
• Isolate response caches by endpoint and bypass cached Jev responses for moving model aliases.
• No collection schema migration is required.

The existing dynamic-field size limit for high-degree entities remains tracked in #51.

Upgrade with uv add "vector-graph-rag==0.3.0".