v0.2.1

• Requires laravel/ai ^0.8.0 (dependency bump; no code changes). Same fix set as v0.2.0, re-tagged on top of the dependency update because the published v0.2.0 reference is immutable on Packagist.

v0.2.0

Fixed — runs always burned their full step budget, even on simple questions:

• New thresholds.goal_satisfied (default 0.70): a goal now counts as satisfied at this verifier score instead of requiring ~1.0, which real embedders/judges almost never return. Applied consistently in the goal scorer (satisfaction snapping), the forward-search frontier (openGoals) and dependency gating.
• The grounded early-stop checks the live frontier (openGoals === []) instead of the unreachable backward_score >= 0.999.
• Backward goal-tree expansion is now guarded: it only refines leaves the best trail has actually worked on and still failed, and is skipped when the remaining budget cannot cover the current frontier anyway — previously the tree grew faster than trails could satisfy it, making termination impossible.
• RecursiveGoalScorer constructor: satisfiedEpsilon replaced by satisfiedThreshold (default 0.7).

Note: superseded by v0.2.1 — the published v0.2.0 reference on Packagist pointed at a tag predating the laravel/ai ^0.8.0 bump.

v0.1.0

Initial release of Laravel BES-RAG — a deep RAG search orchestrator built on the Laravel AI SDK, implementing the search method from Bidirectional Evolutionary Search (BES).

Added

• BES search loop (decompose → seed → evolve → finalize) with budget, LLM call cap and stagnation stop conditions
• Backward decomposition into checkable sub-goals with declarative verifiers (semantic coverage, evidence presence, citation support, entity match, contradiction check, dependency gating)
• Evidence trails as candidates with expand / combine / delete / translocate / crossover operators (goal-boundary aware)
• Boltzmann parent selection with temperature annealing and complementarity-based pair selection
• Bucket-interpolated scoring (groundedness dominates, backward score breaks ties) and a weighted alternative
• Persistence of runs, goal trees, candidates, steps, evidence chunks and goal scores
• Sync engine and queue pipeline (StartRun → SeedCandidates → SearchStep* → FinalizeAnswer)
• Opt-in HTTP API: deep-answer, run status, debug trace, SSE progress stream
• Laravel AI SDK integration (text, structured output, embeddings) with embedding cache
• Testing utilities: FakeLlm, FakeEmbedder, ArrayRetriever
• Per-run retrieval context (->retrievalContext([...])) delivered to the retriever as RetrievalQuery->filters, persisted with the run for the queue pipeline

Full install & usage: see the README.