v0.1.0

Features

• Introduced LENS (Latent Evidence Navigation and Search), a budget-constrained retrieval framework for locating source-grounded evidence directly from raw and dynamic corpora.
• Added a prior-warmed agentic DEEP loop that combines pre-retrieved evidence with iterative ReAct exploration and answer synthesis.
• Added format-agnostic retrieval for corpora with non-semantic filenames, extensionless text files, and large JSON/JSONL shards.
• Added hop-aware search strategies for single-hop, bridge, and comparison questions, including optional second-hop re-search.
• Added configurable LENS components:
  • Batch evidence ranking
  • Multi-arm evidence navigation
  • Adaptive exploration/exploitation mixing
  • Reasoning-chain-guided sampling
  • Statistical stopping decisions
• Added rich search responses containing:
  • Final answer
  • Source-linked evidence
  • Retrieval decisions
  • Token, loop, and file-read telemetry
• Added rich, minimal, context, and json response formats across the Python API, CLI, HTTP API, and MCP server.
• Added a comprehensive ResearchOps benchmark framework with experiment orchestration, checkpoints, retries, artifact registries, control gates, reporting, and Pareto analysis.
• Added HotpotQA evaluation support with frozen sample sets, checksums, dynamic G_n/D_n corpus snapshots, raw-corpus synchronization checks, and stale-index evaluation.
• Added BM25-RAG, Hybrid-RAG, ReAct, LightRAG v1.3.6, SDK, lexical, indexing, and closed-book baseline adapters.
• Added lifecycle and scaling evaluation for setup, indexing, storage, update, and query costs.
• Added LENS ablation profiles for evaluating prior formation, sequential exploration, and knowledge reuse.

Enhancements

• Expanded DEEP retrieval into parallel multi-signal probing across keywords, directory structure, knowledge reuse, specification caches, tree indices, compiled summaries, manifests, and corpus catalogs.
• Overlapped query analysis with retrieval probes to reduce DEEP-mode latency.
• Added evidence-coverage reranking and query-specific file selection.
• Added zero-hit recovery through relaxed keyword decomposition and directory scanning.
• Added bounded ReAct exploration when standard retrieval paths return no usable evidence.
• Improved raw-text evidence extraction with query-centered windows and sibling-file expansion.
• Added matched-line anchoring to preserve precise evidence locations from rga results.
• Added evidence sufficiency tracking with sufficient, partial, and absent states.
• Added a layered answer-resolution pipeline for reasoning leakage, target-type mismatches, comparison questions, and evidence-anchored answer spans.
• Added per-phase LLM usage attribution and richer mechanism telemetry.
• Improved benchmark concurrency controls and per-baseline sample concurrency reporting.
• Added surface-form-tolerant exact-match evaluation and independently calibrated semantic judging.

Refactors

• Changed the default search mode from FAST to DEEP across the Python API, CLI, HTTP API, MCP server, Web UI, Docker examples, and OpenClaw integration.
• Replaced the legacy split DEEP retrieval-and-synthesis path with a unified prior-warmed agentic loop.
• Moved refusal and forced-guess behavior behind an AnswerPolicy interface, separating product behavior from benchmark evaluation semantics.
• Replaced return_context with the unified response_format option; return_context=True remains available as a deprecated compatibility shim.
• Standardized result rendering across local CLI, remote CLI, HTTP, MCP, and Web consumers.
• Standardized baseline outputs through a shared retrieval contract.
• Separated retrieval quality, evidence recall, answer quality, and judge outcomes in benchmark reporting.
• Consolidated benchmark operations under benchmarks/run_benchmark.py with assets, smoke-tune, main, dynamic, ablation, queue, report, and status tasks.
• Moved benchmark-only dependencies into the optional benchmarks dependency group.

Fixes

• Fixed cross-request LLM token and phase accounting under concurrent searches by isolating state with contextvars.
• Fixed paginated-document handling in the prior-warmed agentic loop.
• Fixed DEEP answers containing JSON, code blocks, tool output, reasoning traces, or other prediction pollution.
• Added evidence-grounded recovery for otherwise recoverable loop-path refusals.
• Fixed answer granularity for entity, year, numeric, yes/no, and comparison questions.
• Added controlled rga concurrency, queue timeouts, subprocess cleanup, circuit breaking, and native rg fallback.
• Preserved successful keyword results when another parallel search term fails.
• Fixed retrieval over large raw wiki shards and files with meaningless filenames.
• Fixed dynamic-corpus construction, deterministic title resolution, and snapshot reproducibility.
• Fixed benchmark cache identity to prevent incompatible assets from being reused.
• Fixed propagation of HOTPOT_MAX_CONCURRENT into baseline execution.
• Fixed nested evidence extraction in the shared baseline retrieval contract.
• Fixed semantic judge self-evaluation and clarified judge criteria.
• Prevented judge failures from being silently counted as incorrect answers.
• Added validation for stale-index runs and refreshed generated evaluation tables.
• Made igraph and leidenalg optional at import time; meta-cluster detection is skipped gracefully when unavailable.

Documentation

• Added the LENS paper link: arXiv:2608.16185.
• Reframed the retrieval architecture around budgeted evidence exploration and source-grounded evidence localization.
• Added updated architecture, LENS framework, and knowledge-graph diagrams.
• Added comprehensive English and Chinese ResearchOps benchmark guides.
• Documented frozen evaluation, dynamic corpus construction, baseline lifecycle accounting, ablation studies, and reporting workflows.
• Updated Python, CLI, HTTP API, MCP, Docker, and OpenClaw examples for DEEP mode and response_format.
• Added documented environment templates for benchmark profiles, concurrency controls, and resilient rga execution.

Removed

• Removed the legacy FinanceBench-specific benchmark implementation.
• Removed obsolete experimental scaffolding and superseded benchmark workflows.

Compatibility Notes

• The default search behavior now uses mode="DEEP" instead of mode="FAST".
• The default output is now a rich Markdown evidence report.
• Use response_format="minimal" to retain short-answer behavior.
• Use response_format="context" or "json" for structured telemetry.
• Existing callers using return_context=True remain supported through a deprecated compatibility path.

What's Changed

• [Documentation] Add knowledge graph screenshot to readme by @fanqiNO1 in #172
• [Feature & Refactor] Upgrade framework and add benchmarks by @wangxingjun778 in #175

Full Changelog: v0.0.9...v0.1.0