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]

2.0.0a2 - 2026-10-26

Stored data is not migrated from 2.0.0a1: back up the control database and re-ingest every source. Follow the upgrade guide before upgrading; entries marked Breaking need action.

Added

harborrag

• The harborrag CLI ships with every install; harborrag[local] adds the local-stack clients [#59]

harborrag-adapters

• Parsing for legacy Word .doc and PowerPoint .ppt/.pps files [#85]
• SharePoint connector: crawl sites, libraries, folders or selected files, with allowlists and optional site pages [#90]
• SharePoint device-code sign-in, including for the Docker worker [#90] [#99]
• Jira custom fields are filterable on issues and their attachments [#85]

harborrag-app

• harborrag init scaffolds a ready-to-run project with provider presets and data-source setup [#59]
• harborrag ingest run runs an ingestion in-process, without Temporal [#59]
• harborrag doctor checks the project, configuration and every backing store [#59] [#95]
• harborrag auth keys issues, lists and revokes tenant-bound MCP reader keys [#85] [#89]
• --timeout option for ingest watch, wait and start --wait [#95]
• ingest start takes the connector positionally, like ingest run [#95]
• Pause, resume and cancel actions are recorded in the run’s events [#85]

harborrag-engine

• Placeholder graph nodes for links to sources not yet ingested, replaced once the target is published [#85]

harborrag-mcp-server

• find_entities tool to search whole entities by facet [#85]

harborrag-runtime

• Automatic purge of retired document versions after a configurable retention period [#85]
• Searchable entity summaries that lead to their supporting evidence [#85]
• Summary facets: filterable attributes copied from connector fields [#85]
• Configurable entity summary length [#85]

Fixed

harborrag-adapters

• Corrupt or encrypted .xls files are reported as parser rejections [#85]
• SharePoint handles deleted items and expired tokens without failing the run [#96]
• Legacy secrets are migrated to the correct tenant [#96]

harborrag-app

• Interrupting or failing ingest run no longer leaves the run stuck [#95]
• The API no longer submits the same run twice [#95]
• ingest pause, resume, cancel, status and watch report correct results for finished and partial runs [#95]
• Clearer error messages that never include credentials, and faster CLI startup [#95]

harborrag-engine

• Retrieval works for users who can read more than 10,000 documents [#96]
• Entity search honours all filters, and facet filters are case-insensitive [#96]
• Graph reads and relation repair handle placeholder nodes and cross-scope links reliably [#96]

harborrag-mcp-server

• Tool output schemas are valid for MCP clients again [#85]
• find_entities reports unknown facets instead of returning an empty result [#96]
• A broken API key file denies access instead of revoking every key [#96]
• Over-limit mcp.yaml values are lowered with a warning instead of stopping the server [#96]

harborrag-memory

• Memory retention settings are enforced [#96]
• Memory recall works without an embedding model [#96]

harborrag-runtime

• Large sources no longer time out during discovery [#85]
• Pausing and resuming an ingestion no longer fails the run or loses its progress [#96]
• Re-ingesting content while its old version is being purged retries instead of failing [#96]
• A retried chat request no longer interferes with a slow response still in progress [#96]
• Summary and chat client failures no longer take down ingestion workers [#96]
• A late cancellation can no longer overwrite a completed agent run [#96]
• A warning is logged when a removed storage prefix setting is still set [#96]
• One concise log line per failed document [#95]

Deployment

• Removed unused variables from env-example [#96]

Changed

harborrag-adapters

• Breaking: one storage namespace per tenant across Qdrant, FalkorDB and the object store; re-ingest after upgrading [#85] [#96]
• LiteParse is the default parser, and the OCR server is now opt-in [#85]
• config/parsers.yaml supports environment variables [#85]
• Breaking: SharePoint requires the Sites.Selected permission and limits file types by default [#90]
• SharePoint shares one rate limit across workers, skips unreadable folders, and caps retries per file [#96] [#101]
• Breaking: Jira custom-field filter keys handle accented names; update saved filters and re-ingest [#96]
• Local connectors reject unknown filters, and partial scans no longer retire documents [#95]

harborrag-app

• Breaking: standardized CLI exit codes; failed or partial runs now exit 1 [#95]
• In a repository checkout, ingestion uses the stack’s PostgreSQL and object-store settings instead of a local SQLite database [#95] [#99]
• auth keys requires an up-to-date control database and reports clearer errors [#95]
• Run ids cannot be reused [#95]
• Clearer ingestion API responses when Temporal is unreachable or no longer knows a run [#95]
• Inline progress display replaces the full-screen dashboard [#59]
• Local commands no longer require Temporal to be installed [#59]
• harborrag init keeps existing project files [#96]
• CLI file paths resolve from the current directory [#96]

harborrag-engine

• Each source link is stored as a single graph edge, and duplicate attachment edges are removed [#85]
• Duplicate search results are collapsed per source instead of across the whole corpus [#85]
• Vector search results link directly to their graph nodes [#85]
• More accurate unresolved-relation counts [#85]

harborrag-mcp-server

• Breaking: graph chunk edges now point from structure to chunk, and graph traversal limits are lower [#96]
• The MCP Explorer respects each tenant’s limits [#96]

harborrag-memory

• Project-scope memories stay within the session unless enabled, and summary prompts are bounded [#96]

harborrag-runtime

• Summarization is off by default [#85]
• Summaries indicate when some source documents are still missing [#85]
• Summary reads no longer lock or write data [#85]
• Faster summary updates and publication [#96]
• Retired-version cleanup starts sooner and exposes metrics [#96]
• Ingestion defaults tuned for smaller hosts [#85]
• Faster ingestion workflows, with per-document cost independent of source size [#85]
• The secrets encryption key is only required by processes that open secrets [#85]
• Failed documents report a clear error code and stage, and parser rejections are not retried [#95]
• Startup fails on mismatched embedding dimensions [#95]

Deployment

• MinIO uses a new image that runs as a non-root user [#85]

Deprecated

harborrag-adapters

• The Confluence max_child_pages setting is ignored with a warning [#96]

Removed

harborrag-app

• Textual ingestion dashboard [#59]

harborrag-mcp-server

• Breaking: composed_evidence_search, verify_citations and get_document_metadata tools; use vector_search, fetch_evidence and get_document_context instead [#85]

harborrag-memory

• Unused HARBORRAG_MEMORY_PII_REDACTION setting [#96]

Security

harborrag-adapters

• The .ppt parser is protected against crafted files that expand into huge output [#96]
• SharePoint tokens are stored safely, and pre-authenticated download links are no longer kept [#96] [#98]
• Model configuration errors no longer print API keys [#98]

harborrag-app

• Project discovery only trusts directories owned by the current user [#59]
• Errors and logs no longer expose credentials [#95]
• Unauthenticated mode refuses non-local requests unless explicitly allowed [#98]

harborrag-core

• More credential field names are detected and redacted [#98]

harborrag-engine

• Retrieval diagnostics no longer reveal documents the caller cannot read [#96]

harborrag-mcp-server

• Reader keys cannot target all tenants [#85]
• In API key mode, the owner token is only accepted for single-tenant deployments [#96]
• The MCP server no longer receives the secrets encryption key [#96]

harborrag-memory

• Shared memories can only be deleted or replaced by their author or a user with the required role [#96]

harborrag-runtime

• A tenant reindex no longer touches other tenants’ cleanup jobs [#98]

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.

Added

harborrag

• HarborRAG 2.0: a modular, provider-agnostic RAG framework built from eight independently testable packages [#1] [#6]
• One harborrag install, with extras for the CLI, HTTP API, MCP server, chat, Temporal and each storage provider [#31]
• Python SDK with ingestion, retrieval, graph and chat services behind one async client [#32]

harborrag-adapters

• Connectors for local files, GitHub, Confluence, Jira and SharePoint, including comments and attachments [#2] [#6]
• Parsers for text, Markdown, HTML, JSON, CSV/Excel, Word, ODT, EPUB, PowerPoint, images and PDF [#2] [#9]
• Choice of PDF engines: PyMuPDF, Docling, LiteParse, MinerU or PaddleOCR [#2] [#9]
• Chat, embedding and reranking clients with configurable providers [#4]
• Storage on PostgreSQL, S3-compatible object storage, Qdrant, FalkorDB and Redis [#5] [#7]
• Confluence pages are converted into structured text and tables [#16]
• Optional attachment ingestion for Confluence and Jira [#19]
• Local file discovery reports skipped files with a reason [#14]

harborrag-app

• harborrag operator CLI [#1]
• HTTP control-plane API with authentication, role-based access, health checks and a standard error format [#7]
• Read APIs for projects, sources, activity, settings and metrics [#11]
• Ingestion and connection APIs with live progress streaming [#25] [#34]

harborrag-engine

• Ingestion pipeline that skips unchanged documents instead of re-parsing and re-embedding them [#12] [#22]
• Table-aware chunking [#15] [#16]
• Dense, sparse and hybrid retrieval that only cites the current version of a document [#12] [#27]
• Knowledge graph of document structure and source relationships [#12] [#38]
• Retrieval-grounded chat with streaming, and a bounded multi-turn agent [#22]

harborrag-mcp-server

• MCP server with vector_search and graph tools for triplets, paths and subgraphs [#10] [#22] [#27]

harborrag-memory

• Short-term and working memory for chat sessions [#22]

harborrag-runtime

• Durable ingestion on Temporal: survives worker restarts, retries single documents, and supports pause, resume and cancel [#12] [#24]
• Rebuild vector and graph indexes from stored content without re-crawling sources [#12]
• Temporal settings in YAML configuration [#32]

Deployment

• Docker Compose stack and dev.sh scripts to run the full stack locally [#22] [#32]

Documentation

• Public documentation website with interactive architecture and feature sections [#31] [#40]
• Data-lifecycle, runtime-reliability, publication and security guides [#31]

Development

• Pre-commit and pre-push hooks running the same quality gates as CI [#28]

Fixed

harborrag-adapters

• CMYK images are converted before OCR [#37]
• Clearer errors for undecodable text files and password-protected presentations [#18] [#22]

harborrag-engine

• Retrieval never returns a superseded document version during a concurrent publish [#31]
• Graph subgraph search returns complete results and no longer times out on large graphs [#35]
• Jira subtasks no longer appear as a separate node from their parent issue [#33]
• The graph keeps page and issue membership separate from document structure across all connectors [#38]
• Relation filters only accept relation types the graph actually stores [#36]

harborrag-mcp-server

• The local MCP HTTP server works on Windows [#26]

harborrag-runtime

• Parsing falls back to in-process when a subprocess cannot start, instead of failing the document [#30]

Documentation

• Website build fixes for link checking, coverage pages, UTF-8 output and asset paths [#43] [#44] [#45] [#48]

Changed

harborrag

• License changed to Apache 2.0 [#3]

Security

harborrag-adapters

• Connector credentials are encrypted at rest [#25]
• Outbound requests to private or disallowed addresses are blocked [#8]

harborrag-app

• Every control-plane API call is limited to the caller’s tenant [#23]
• Secrets are redacted from API responses and diagnostics [#8] [#11]
• Production refuses to run on the default SQLite database [#8]

harborrag-mcp-server

• The audit log refuses symlinks and other unsafe files [#26] [#27]

Deployment

• Database, MinIO and Grafana passwords must be supplied by the operator [#31]

This site is open source. Improve this page.