Langroid Files

Batch processing

Two fixes to run_batch_* that affected correctness of concurrent and cancelled batches:

• Exception policies now apply to output_map and to cancellation (#1144). output_map previously ran outside the exception handler, so a mapper error aborted the whole batch even under RETURN_NONE / RETURN_EXCEPTION, and error placeholders were handed to the mapper. Cancellation was caught as BaseException and recorded as a normal result, and sequential batches kept starting new items after the caller had cancelled. Additionally, an exception object returned by a task is no longer confused with a failed task, and a one-shot iterator input under RAISE now raises instead of silently yielding []. Cancellation semantics on Python 3.10 vs 3.11+ are documented in docs/notes/batch-processing.md. Supersedes [#1140] and [#1142] by @feng1201, both cherry-picked with authorship preserved — thank you.

• Agent copies no longer share a mutable config (#1141, by @feng1201). run_batch_agent_method reused one configuration object for every agent copy, so names accumulated suffixes (Worker-0-1-2) and concurrent methods could observe another copy's name or mutable config state across an await. Each item now gets a deep copy, with its name derived from the original agent name and the global item index.

Retrieval and search

• Fuzzy search keeps documents attached to their own scores (#1145, by @dingpuyu). When one cleaned chunk was a substring of another, the substring lookup mapped both matches to the first document, and reciprocal rank fusion then dropped the second one. The document index from process.extract is now carried through, so each score stays with the right content and metadata.

• Overlapping context windows merge completely (#1146, by @YaoxinHuang). When a later window joined two existing groups, components only relabelled its immediate neighbours, leaving already-connected windows in separate groups — so VectorStore.remove_overlaps could return duplicate chunks. All members of the joined groups are now relabelled.

Documentation

• Added the Cross-Encoder Reranking note, which the docs nav had linked to but which had never existed in the repo, and linked the existing quiet-mode note from the nav (#1143, by @KunyangZhang).
• Fixed a dead docs/development/github-cli.md link in CONTRIBUTING.md that 404'd for first-time contributors at exactly the point where they are being walked through opening a PR (#1148, by @tonydzi).

Full Changelog: https://github.com/langroid/langroid/compare/0.68.0...0.68.1