HarborRAG

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Fixed

• Revalidate retrieval candidates against the authoritative active document versions immediately before returning results, preventing a concurrent publication from exposing a superseded document version.

Security

• Require operator-supplied PostgreSQL, MinIO, and Grafana passwords instead of checked-in Compose fallbacks. Existing development environments must add the required database credentials and create the separate protected monitoring environment file before restarting the affected stacks.

2.0.0a1 - 2026-08-27

This is the first public HarborRAG 2.0 alpha release. Python package metadata, distribution filenames, dependency pins, and release tags use the canonical PEP 440 version 2.0.0a1; release tooling also accepts 2.0.0-alpha as an operator-facing alias.

Documentation

• Rebuilt the public documentation website around HarborRAG 2.0.0 capabilities, limitations, package boundaries, and project lineage.
• Added a publication-manifest hero, interactive ingestion/retrieval explorer, accessible interface examples, and responsive navigation while retaining the existing visual assets and splitting presentation code into focused modules.
• Added public data-lifecycle, runtime-reliability, publication, and security guidance plus a CI guard against publishing private reference material.

Changed

Project renamed: Qdrant-loader → HarborRAG

• Qdrant-loader is now HarborRAG, restructured from a single Qdrant-coupled ingestion tool into a modular, provider-agnostic RAG framework. The old qdrant-loader / qdrant-loader-core / qdrant-loader-mcp-server packages are replaced by eight ports-and-adapters packages (harborrag-core, harborrag-adapters, harborrag-engine, harborrag-memory, harborrag-runtime, harborrag-app, harborrag-mcp-server, harborrag), so connectors, parsers, model providers, and storage backends can evolve independently instead of being hardwired to Qdrant and a fixed provider set.
• Documentation, Docker assets, and CI workflows updated to reflect the new package layout and naming.

This is a breaking rename: import paths, package names, and CLI/config references from qdrant-loader do not carry over. See README.md for the current architecture.

1.0.3 - 2026-06-08

Added

Qdrant-loader

• Queue-based ingestion architecture with SQLite job queue, worker pool, scheduler, and new serve/jobs/webhook CLI flows [#269] [#280] [#281] [#284] [#289] [#291] [#294] [#307]
• Excel ingestion improvements and Jira Data Center custom field mapping support [#251] [#276]
• updated_after filter for Jira connector to limit ingestion to issues updated after a given date [#267]
• Container deployment setup with dedicated Dockerfiles and docker-compose.yaml [#278]

Qdrant-loader-core

• Bedrock and Gemini provider support in the LLM/embedding layer [#275] [#292]
• Sparse retrieval primitives for hybrid search flows [#184]

Qdrant-loader-mcp-server

• Qdrant hybrid search support with sparse+dense pipeline integration [#184]

Fixed

Qdrant-loader

• Queue concurrency and failure-path hardening, including race-condition fixes and safer startup/auth/readiness handling [#307] [#311]
• State DB and Alembic path resolution reliability, plus safer rollback behavior for multi-project transactions [#270] [#285] [#290]
• Deleted Jira issues are now cleaned up from Qdrant during re-ingestion and delete flows [#307]
• Empty-chunk regression prevention and CORS environment configuration fixes [#268] [#293]

Changed

Qdrant-loader

• Ingestion pipeline refactored to a streaming-oriented flow for queue-driven execution [#295]

1.0.2 - 2026-05-12

Changed

Documentation

• Updated contributing guide for external contributors [#273]

1.0.1 - 2026-04-23

Fixed

Qdrant-loader

• Fixed UI layout issues across website pages [#253]
• Privacy policy last_updated fallback now uses template file mtime instead of build date when Git timestamp is unavailable [#259]

Documentation

• Updated docs and removed broken internal links [#258]

1.0.0 - 2026-04-09

Added

Qdrant-loader

• Contextual embeddings for enriched chunk context during ingestion [#221]

Fixed

Qdrant-loader

• Jira Cloud connection failure due to deprecated search API endpoints [#215]
• Duplicate chunks for Python files by rewriting AST parser [#217]
• Duplicate document IDs causing missing chunks in metric tracking [#222]
• Ingestion metrics: aligned size metrics and aggregated project results [#222]
• JQL injection and query breaking when configuration values include special characters [#223]
• Jira connector validation for environment configuration and connection errors [#224]
• Optional issue_types and include_statuses configuration for Jira connector [#223]
• Error message sanitization to prevent accidental exposure of sensitive data [#230]
• Contextual embedding content formatting [#228]
• YAML syntax error and broken markdown in configuration documentation [#231]

Qdrant-loader-core

• Sensitive data redaction in logging processors [#230]

Qdrant-loader-mcp-server

• Embedding logic and configuration for local (Ollama) model support [#229]
• vector_search_service to use correct embedding_model parameter [#229]
• Retry logic and failure context reporting in search pipeline [#229]

Changed

Qdrant-loader

• Aligned configuration templates with Ollama-first and OpenAI unified setup [#231]
• Migrated pip/venv to uv for monorepo package management [#220]
• Updated deprecated GitHub Actions from Node.js 20 to Node.js 24 [#227]
• Updated website UI to v1 [#219]

0.9.0 - 2026-03-26

Added

Qdrant-loader

• enable_semantic_analysis global NLP kill switch in chunking config to skip spaCy/LDA processing entirely for faster ingestion [#189]
• enable_enhanced_semantic_analysis opt-in flag in chunking config to gate advanced NLP fields (pos_tags, dependencies, document_similarity) [#195]

Qdrant-loader-mcp-server

• expand_chunk_context MCP tool to retrieve surrounding chunks for richer context around a specific chunk [#185]
• cluster_session_id required parameter to expand_cluster tool schema for session isolation [#175]

Fixed

Qdrant-loader

• Semantic analyzer now filters out garbage/punctuation-only entities from NLP metadata [#178]
• Improved semantic similarity caching to avoid self-matches and keep similarity results fresh across repeated enhanced analyses [#195]

Qdrant-loader-mcp-server

• Instance-level _clustering_cache causing cross-request data leakage in IntelligenceHandler; replaced with session-scoped store with TTL cleanup [#175]
• MCP server concurrency improved with FastAPI-centric workers and semaphore-based search load distribution [#176]
• Removed redundant top-level schemas.py that duplicated per-tool schema modules [#181]
• Search handler config instance de-duplicated to eliminate redundant initialization [#174]

Changed

Qdrant-loader-mcp-server

• CLI and transport layer refactored: server.py centralizes startup; HTTP handler split into focused route modules [#176]

0.8.1 - 2026-03-24

Fixed

• CI secret injection restricted to develop/main branches only [#187]
• CI secrets and integration tests skipped for fork/Dependabot PRs [#187]

Changed

• Removed uv.lock from version control tracking [#187]

0.8.0 - 2026-03-17

Removed

Qdrant-loader

• Legacy binary Office formats (.doc/.ppt) no longer supported [#145]

Fixed

Qdrant-loader

• Unsupported-type fallback now skips ingestion for legacy .doc/.ppt files [#145]
• Excel chunk overlap consistency [#142]
• .xls inputs now treated as Excel files for chunking strategy and metadata [#141]

Qdrant-loader-mcp-server

• Applied NLP processing only to keyword-search flow [#134]
• Lazily loaded qdrant-client, spaCy, and network imports in MCP server [#132]
• Reduced CLI startup time by lazily loading the OpenAI SDK [#131]

Added

Qdrant-loader-mcp-server

• Cross-encoder reranking to improve retrieval relevance [#151]
• expand_document MCP tool to expand context around a specific chunk [#149]
• expand_cluster MCP tool with detailed analysis of document grouping and relationships [#146]
• WRRF ranking for hybrid search to improve result quality [#133]

Changed

• Simplified configuration with smart defaults and a setup wizard [#144]

0.7.6 - 2026-01-22

Changed

• Documentation terminology updated from “Release Notes” to “Changelog” [#112]
• Applied consistent formatting across workflow and updated navigation links [#112]

0.7.5 - 2026-01-20

Fixed

Qdrant-loader-mcp-server

• Mixed document types (dict/object) causing “Untitled” and empty content_preview in results [#92]
• Incorrect usage of NestedCondition for project_ids filter replaced with dot notation [#97]
• Search results not respecting user-specified limit parameter (now defaults to 5) [#98]

Added

Qdrant-loader-mcp-server

• Optional similarity_threshold parameter (default 0.7) to find_similar_documents for filtering by minimum similarity score [#91]
• Separate fetch_limit from user limit with configurable defaults for better control [#98]

Changed

Qdrant-loader-mcp-server

• Updated reason field from explanation to recommendation_reason in complementary content results [#100]
• Enhanced validation for text field types to prevent display issues [#92]

0.7.4 - 2025-12-11

Fixed

Qdrant-loader

• Windows asyncio event loop crashes with signal handling and stdio support [#76]
• –log-level CLI option not working after logging initialization [#63]
• Missing prometheus-client dependency causing import errors [#75]
• Deprecated langchain.text_splitter import warnings [#76]

Qdrant-loader-mcp-server

• Missing Spacy dependency causing Cursor MCP crash on startup [#82]
• MCP Search API compatibility issues with qdrant-client 1.16 [#78]

Added

Qdrant-loader

• Test suite cross-platform support [#76]

0.7.3 - 2025-09-11

Fixed

Qdrant-loader

• Logging duplication where CLI commands printed each log message 2-4 times [#56]

Changed

Qdrant-loader-core

• Unified logging architecture with centralized configuration across all packages [#56]
• Idempotent setup with new reconfigure() method [#56]

0.7.2 - 2025-09-05

Changed

• Internal inter-package dependencies now pinned to unified release version [#54]
• Enhanced dry-run output to preview internal dependency pin changes [#54]
• Enforced explicit release order: bump → update classifiers → pin deps → commit → tag → release [#54]
• Updated commit messages to reflect classifier and internal dependency updates [#54]

0.7.1 - 2025-09-04

Added

Qdrant-loader-core

• Azure OpenAI support with robust endpoint handling (BETA) [#52]
• Ollama endpoint handling [#52]
• Unified global.llm.* configuration for provider-agnostic setup [#52]
• Structured logging for LLM requests (provider, operation, model, latency) [#52]
• Secret redaction in logs [#52]
• Normalized exception mapping across providers [#52]

Changed

Qdrant-loader-core

• Vector size now read from config instead of hardcoded 1536 defaults [#52]
• Centralized LLM layer with provider adapters [#52]
• Direct OpenAI imports removed from application code [#52]
• Tests and documentation updated for provider-agnostic architecture [#52]

Qdrant-loader-mcp-server

• MCP server prefers config file loading with CLI/env/file precedence [#53]

Deprecated

Qdrant-loader-core

• Legacy fields global.embedding.* and file_conversion.markitdown.* (deprecation warnings shown) [#52]

Security

Qdrant-loader-mcp-server

• Redacted --print-config for MCP server [#53]

0.6.1 - 2025-08-13

Fixed

Qdrant-loader-mcp-server

• Timeout and error issues in detect_document_conflicts tool (achieved P95 latency 8-10s) [#45]
• AttributeError in conflict formatter: 'dict' object has no attribute 'document_id' [#45]

Added

Qdrant-loader-mcp-server

• Tiered analysis with intelligent document pair prioritization [#45]
• Parallel processing with concurrent Qdrant vector retrieval [#45]
• 8 configuration options for conflict detection performance tuning [#45]
• Detailed performance statistics in tool responses [#45]
• Optional per-call parameter overrides for conflict detection [#45]

Changed

Qdrant-loader-mcp-server

• Strict budgeting for expensive LLM calls (default 2 pairs) [#45]
• Graceful degradation with partial results instead of hard failures [#45]
• Tool schema enhanced with performance parameters [#45]
• Error handling for malformed document data [#45]

0.6.0 - 2025-08-12

Added

Qdrant-loader-mcp-server

• FastAPI-based HTTP transport alongside stdio transport [#43]
• Server-Sent Events (SSE) streaming capabilities [#43]
• --transport CLI option to choose between stdio and HTTP modes [#43]
• Health check endpoints for production deployment [#43]
• Structured tool output with JSON content [#43]
• Tool behavioral annotations for all 8 tools [#43]
• Protocol version validation with graceful degradation [#43]
• Comprehensive session management for HTTP connections [#43]

Changed

Qdrant-loader-mcp-server

• MCP Protocol upgraded from 2024-11-05 to 2025-06-18 [#43]
• Modular transport layer with clean separation [#43]
• Error handling with improved messages [#43]
• Connection handling performance optimizations [#43]

0.5.1 - 2025-07-28

Changed

Qdrant-loader-core

• All chunking strategies refactored into modular components [#39]
• Dedicated classes for document parsing, section splitting, metadata extraction [#39]
• HTML chunking with robust handling for empty content and malformed HTML [#39]
• Code and JSON strategy complete modular redesign [#39]
• Chunk processing with additional metadata fields [#39]
• Configuration templates with new strategy-specific options [#39]

0.5.0 - 2025-07-25

Added

Qdrant-loader-mcp-server

• Cross-document intelligence: similarity analysis, clustering, relationship detection [#35]
• Intent-aware adaptive search with AI-powered query understanding [#35]
• Knowledge graph integration with entity relationships [#35]
• Topic-driven search chaining with automatic discovery [#35]
• Dynamic faceted search with real-time generation [#35]
• spaCy integration for advanced NLP processing [#35]

Qdrant-loader

• --force flag to bypass change detection for complete reprocessing [#35]

Changed

Qdrant-loader-core

• Topic extraction with enhanced LDA modeling [#35]
• Entity recognition with structured conversion [#35]
• Chunking with improved timeout handling [#35]
• Structured logging for semantic analysis [#35]

0.4.15 - 2025-07-22

Fixed

Qdrant-loader-core

• Critical chunking inconsistency (all strategies now use character-based chunk_size) [#33]

Changed

Qdrant-loader-core

• Markdown strategy refactored into focused components [#33]
• Hierarchical metadata with intelligent section analysis [#33]
• Split level detection based on document structure [#33]
• Boundary detection using tokenizer for word/token boundaries [#33]

Added

Qdrant-loader-core

• Comprehensive integration tests for strategy consistency [#33]

0.4.14 - 2025-07-13

Fixed

Qdrant-loader-core

• Regex error in Excel table detection: bad character range |-\s [#33]
• Large Excel tables treated as single massive chunks [#33]
• Token limit warnings for large Excel chunks [#33]

Changed

Qdrant-loader-core

• Logical unit management with intelligent splitting at line boundaries [#33]
• Efficient chunking for large Excel files [#33]
• Error handling with better error messages [#33]
• Table structure preservation during chunking [#33]

0.4.13 - 2025-07-11

Added

Qdrant-loader-core

• Sheet-aware sectioning for Excel files (split on H2 headers) [#33]
• Table-aware chunking with specialized _split_excel_sheet_content method [#33]
• Intelligent content detection based on original_file_type metadata [#33]
• 3 test cases for Excel chunking scenarios [#33]

Changed

Qdrant-loader-core

• Excel-to-markdown chunking in MarkdownChunkingStrategy [#33]
• Context-aware splitting with different header level thresholds [#33]
• Metadata tracking with is_excel_sheet field [#33]
• Table boundary preservation in chunking [#33]

0.4.12 - 2025-07-10

Fixed

Qdrant-loader-core

• Missing chunk overlap in MarkdownChunkingStrategy [#33]

Added

Qdrant-loader-core

• Intelligent overlap calculation using paragraph/sentence boundaries [#33]
• Comprehensive overlap testing [#33]

Changed

Qdrant-loader-core

• Overlap configuration support (0 for no overlap, up to 25% for context) [1f53556]

0.4.11 - 2025-07-10

Fixed

Qdrant-loader-core

• File size detection limits for larger documents [6160435]
• MarkdownChunkingStrategy not respecting chunk_size configuration [6160435]
• Unique chunk ID generation (chunks had identical IDs causing overwrites) [6160435]

Added

Qdrant-loader-core

• max_chunks_per_document configuration parameter [6160435]

Changed

Qdrant-loader-core

• Chunk count management with configurable limits [6160435]
• Error messages with actionable configuration advice [6160435]
• Section limits now dynamic (50% of max_chunks_per_document) [6160435]

Removed

Qdrant-loader-core

• Conflicting max_document_size parameter [6160435]

0.4.10 - 2025-06-18

Fixed

Qdrant-loader

• Duplicate debug logging with [DEBUG] [DEBUG] tags [c72a872]
• Mixed path separators in Windows log output [c72a872]
• .txt file processing when file_types: [] was empty [c72a872]

Qdrant-loader-core

• Windows file URL parsing for LocalFile connector [c72a872]
• Git connector document URL generation with Windows paths [c72a872]

Changed

Qdrant-loader

• Logging verbosity control for third-party libraries [c72a872]
• Emoji handling for Windows console [c72a872]
• SQLite logs now suppressed [c72a872]

Qdrant-loader-core

• Path normalization across all connectors [c72a872]
• Timeout handling (threading on Windows, signals on Unix) [c72a872]

Added

Qdrant-loader

• 38 Windows compatibility test cases [c72a872]

0.4.9 - 2025-06-18

Fixed

Qdrant-loader-core

• Missing content_type="md" field in _create_deleted_document method [430cb2a]

0.4.8 - 2025-06-17

Fixed

Qdrant-loader-core

• Windows file URL parsing in LocalFile Connector [b60e6e0]
• Git Connector document URL generation with Windows paths [b60e6e0]
• File conversion timeout handling for cross-platform [b60e6e0]
• MarkItDown Windows signal compatibility [b60e6e0]

Qdrant-loader

• Console emoji handling for Windows [b60e6e0]
• Duplicate log level display [b60e6e0]

Added

Qdrant-loader

• 38 Windows compatibility test cases [b60e6e0]

0.4.7 - 2025-06-09

Fixed

Qdrant-loader

• Upgrade instructions to include qdrant-loader-mcp-server package [fecc0dc]

Qdrant-loader-core

• Branch display logic to default to ‘main’ when unknown [fecc0dc]

Changed

Qdrant-loader

• Configuration template with detailed comments [fecc0dc]
• Release script with automatic RELEASE_NOTES.md validation [fecc0dc]

Qdrant-loader-core

• Logging in PublicDocsConnector [fecc0dc]

0.4.6 - 2025-06-03

Added

Qdrant-loader

• Automatic update notifications when new versions available [3604f92]
• Non-intrusive background version checking [3604f92]

0.4.5 - 2025-06-03

Fixed

Qdrant-loader

• Version detection using importlib.metadata.version() [#25]

Qdrant-loader-core

• Circular imports in config → connectors → config cycle [#25]

Changed

Qdrant-loader

• CLI startup time reduced by 60-67% for basic commands [#24]
  • --help: ~6.8s → 2.33s (66% improvement)
  • --version: ~6.3s → 2.57s (59% improvement)
• Lazy loading for heavy modules (96-97% import time reduction) [#25]

Added

Qdrant-loader-core

• Warning capture system for Excel file processing [#25]
• Structured logging for openpyxl warnings [#25]
• Smart detection for “Data Validation” and “Conditional Formatting” warnings [#25]
• Summary reporting for unsupported Excel features [#25]

0.4.4 - 2025-06-03

Fixed

Qdrant-loader-core

• File conversion initialization with missing set_file_conversion_config calls [#21]
• Converted files using wrong chunking strategy [#21]
• NLP processing skipped for converted files [#21]
• MarkdownChunkingStrategy infinite loops with very long words [#21]

Qdrant-loader

• ResourceManager cleanup causing workers to exit prematurely [#21]

Added

Qdrant-loader-core

• Safety limits: MAX_CHUNKS_PER_SECTION = 100 and MAX_CHUNKS_PER_DOCUMENT = 500 [#21]
• Handling for words longer than max_size [#21]
• Comprehensive tests for converted file NLP processing [#21]

Changed

Qdrant-loader-core

• Strategy selection based on conversion status [#21]
• Metadata propagation for converted files [#21]
• MAX_CHUNKS_TO_PROCESS increased from 100 to 1000 chunks [#21]
• Large document handling (up to ~1000KB text limit) [#21]

Qdrant-loader

• Change detection for incremental updates [#21]
• Signal handling for graceful shutdown [#21]

This site is open source. Improve this page.