v0.4

Added

• Optional experiment loading modes on ENCODE: load_mode="eager" (default, backward compatible), load_mode="lazy", and load_mode="incremental".
• Incremental summary loading with load_experiment_summaries(), get_experiment_summaries(), and load_next_experiment_batch() to keep lightweight summaries in memory until full experiment payloads are explicitly materialized.
• Summary cache stored separately in experiment_summaries.json to support incremental and indexed search flows.
• Optional search indexing with build_search_index(), clear_search_index(), and get_search_index_stats().
• Batch search helper: search_experiments_batch() for grouped biosample, organism, and target searches.
• Facet helper: get_experiment_facets() for counts by assay, biosample, organism, status, and target.
• Export helper: export_experiments() for JSON, CSV, and TSV output.
• Performance reporting: get_performance_stats() with lightweight HTTP/cache/index metrics.
• Batch file-accession helpers on ENCODE: get_file_metadata_batch(), get_file_url_batch(), and search_experiments_by_file_accessions().
• Download manifest and verification support in encodeExperiment: prepare_download_manifest(), preview_only, resume, and verify_checksum options.
• New MCP tools:
  • search_batch
  • get_experiment_facets
  • rebuild_search_index
  • get_search_index_stats
  • get_performance_stats
  • export_experiments
  • get_file_metadata_batch_by_accession
  • get_file_url_batch_by_accession
  • search_by_file_accession_batch
• Streamlit client quick actions for batch search, facets, exports, and metrics/index inspection.
• Streamlit client LLM provider selection with configurable base URLs for both Ollama and OpenAI-compatible endpoints such as LM Studio.
• Focused unit-test coverage for lazy/incremental loading, indexed search, download preview/resume/checksum behavior, server auth/cache semantics, and server load-mode configuration.

Changed

• Library version bumped to 0.4.
• MCP server version bumped to 0.4.
• Streamlit client version bumped to 0.4.
• Streamlit client now supports session-scoped API keys for auth-enabled MCP servers and injects them automatically into tool calls.
• Streamlit client can now route model requests to either Ollama (/api/chat) or OpenAI-compatible chat APIs (/chat/completions) based on sidebar configuration.
• Search methods now reuse create_experiment_object() when experiment data is already available, avoiding unnecessary object reconstruction by accession.
• list_experiments on the MCP server now uses experiment summaries so low-memory server modes do not have to fully materialize all experiments just to list them.
• download_files() now returns a manifest alongside download results.
• _request_with_retry() now accepts optional request headers, enabling resumable downloads.

Fixed

• MCP server API-key authentication is now actually enforced across all exposed tools when auth is enabled.
• MCP server cache clearing now resets the shared in-process ENCODE singleton so subsequent requests do not reuse stale state.
• clear_cache(clear_metadata=True) on the MCP server now clears both the main experiments cache and metadata cache, matching the user-facing message.
• Metadata cache read/write failures now log warnings instead of failing silently.
• Experiment and file API responses now receive basic structure validation before they are cached or returned.
• Streamlit tool-call displays and saved assistant tool-call payloads now redact API keys instead of exposing them in the UI.

Notes

• ENCODE() with no new arguments remains eager-loading by default for backward compatibility.
• In incremental mode, direct access to encode.experiments still materializes the full experiment list to preserve the existing public API; the low-memory gains come from summary-based search and explicit batch materialization methods.

v0.2

initial release