carrel

English | 中文

• Text queries. An optional retrieval profile ranks vector and keyword matches first. The service then follows graph references, checks each passage against its indexed version, and appends up to six source passages. The response identifies which passages came from graph follow-up.
• Image queries. Text and visual hits share the same top-K. When graph exploration is enabled, queries without a document filter can also return up to two subjects and two one-hop neighbors per subject for further exploration.
• Follow-up evidence. The search Skill carries document and graph versions and the original time scope into follow-ups. A stale reference returns HTTP 409 so the agent can look up the source again. Image and crop requests honor the same time scope. Follow-up suggestions cover specific questions the answer has left open.

The general benchmarks repeat the 0.4.0 text, image and document-answer tests and separately replay the 0.3.0 retrieval suite. The codebase evaluation compares Haiku’s answers to 24 questions with and without follow-up on 0.4.1. Full evaluation

Upgrade from 0.4.0

Update to tag v0.4.1, refresh the bundled carrel-search Skill in your agent, and restart the search service. To enable the text profile used in this evaluation, set KB_SEARCH_TEXT_DISCOVERY=true in the search service environment before restarting. It is opt-in. Existing chunks, indexes and graphs can be used directly.

中文说明

This release connects source-preserving parsing, graph evidence and ontology review more closely. It adds scoped correction APIs and graph-version comparison, and moves multimodal parsing to models selected per knowledge base.

Parsing and evidence

• Preserve code whitespace, line ranges and configuration fields through parsing, chunking, graph extraction and source readback. Resolve code imports within their project and package scope.
• Improve PDF code recognition, section attribution, short-table recovery and row alignment. Recover visible form fields and retain images embedded in spreadsheets with their sheet and row locations.
• Pack adjacent short text when section and source boundaries support it. Keep code, tables and figures attached to their original evidence regions.
• Retain informative visual content, mark truncated or uncertain model output, and preserve original table text when a replacement cannot be aligned.

Graphs and ontology

• Preserve negated, conditional and uncertain claims with their qualifiers. Carry entity identity and scope information into extraction, and validate relationship endpoints against the schema.
• Carry observation dates, units, conditions and source locations into facts and cross-document series. Default retrieval retains historical, conflicting and future claims; fact queries with an explicit cutoff date filter both facts and change summaries.
• Reuse schema definitions supported by current source evidence, account for uncovered document types and retain endpoint-use statistics for review. Corpus changes can introduce new types without treating past frequency as a permanent constraint.
• Compare graph versions and distinguish changed source material from changed extraction output. Keep three graph versions by default and retain schema versions referenced by retained builds.
• Add preview, commit and revoke APIs for corrections scoped to knowledge-base identity, graph version and source evidence. Track stale records and preserve correction outcomes when a build resumes.

Retrieval and model configuration

• Share model-parameter compatibility across connectivity checks, graph calls and image descriptions. Keep free-form model registration and add scoped overrides for custom deployments.
• Improve fused text/image candidate selection, source diversity and bounded original-text excerpts. The search Skill can navigate documents, locate phrases and follow graph evidence with reusable source references.
• Choose one multimodal model from the console registry for each knowledge base. Images, charts and table review share it; later parsing tasks use the current selection.
• Remove the bundled local VLM container. The optional local-model profile now provides four embedding and reranking services. Vision endpoints can run on another host or through a compatible provider.

Web console

• Complete bilingual progress and error messages, including messages containing Chinese model or knowledge-base names.
• Unify parsing and graph-build stages, actual work counts and waiting states. Task details show models, versions and stage records; overall progress is marked as an estimate.
• Add explicit pause and resume controls for parsing.
• Improve file-list layout with stable column widths, a resizable sidebar, compact parsing notes and direct page selection.

Design

Carrel keeps original evidence, RAG indexes and ontology graphs connected. Technical checks establish source identity, structural consistency and time scope. Reviewed corrections capture business interpretation with supporting evidence.

The Agent Loop plugin remains in development and is not yet implemented. It is intended to help agents learn an enterprise ontology, seek clarification from people and contribute evidence-backed governance feedback.

Capability evaluation

The 0.4.0 evaluation cov

English | 中文

This release expands file format support and improves source fidelity, console loading, and macOS iCloud synchronization.

Parsing and source references

• Add static readers for R, Jupyter Notebook, RST, Quarto/RMarkdown, TeX/BibTeX, JSONL, TSV, Nextflow/CWL/WDL, Snakefiles, Feature, DOT, XSD, logs, and recognized extensionless scripts.
• Preserve notebook cells, code and saved outputs together. Carry cell, output, line and row locations through to search results.
• Recover PDF titles and body text misclassified as headers, and retain the original page range of merged tables in citations.
• Preserve TeX section hierarchy and HTML template content. Filter HTML comments from Markdown prose while retaining code examples and front matter.
• Keep standalone image content, avoid stacking generated figure descriptions, and reduce false heading-only chunk warnings.

Ingestion and console

• Filter Git LFS pointers during scanning so they stay out of file lists, counts and parsing queues. Restored originals enter the normal ingestion flow; existing index entries are withdrawn when a file becomes a pointer.
• Fix saving the configuration of a newly enabled knowledge base and keep completed label-extraction tasks in their final state.
• Improve file-list loading on slow connections with compressed responses, request timeouts and recovery. Prevent duplicate requests and results from a previously selected library, and make mobile navigation more compact.

macOS iCloud sync example

• Use a default window of 64 pending downloads, refilling slots as downloads finish. Improve timeouts, retries and fallback from the download helper to the system downloader.
• Recover scan notifications after interrupted transfers and refresh mirror locks during long copies. Updating the sync installation waits for existing agents to stop.

Documentation

• Add the bilingual 0.3.0 retrieval evaluation, covering retrieval accuracy, graph contribution and parallel reranking, with links from the README and retrieval guide.

Upgrade from 0.3.0

1. Update the checkout to v0.3.1 and run ./deploy.sh with your usual deployment options.
2. On Linux, run systemctl --user restart carrel-web carrel-search. On other hosts, restart the corresponding application processes.
3. If you use the macOS iCloud sync example, update the scripts in your installed sync folder and rerun ./install.sh, keeping your existing sync.env.

Existing documents retain their parsed results and caches. Newly supported formats enter through incremental scans and follow each library's graph settings. Reparse selected files in the console to apply the parsing improvements to previously indexed documents.

For upgrades from 0.2, first follow the graph migration steps in the 0.3.0 release notes.

English | 中文

Changes

• Versioned LinkML schemas with eight upper classes, type inheritance, class/instance distinctions, and scoped entity identity.
• Table row extraction, external identifiers, source references, and current/historical evidence views.
• Automatic label sampling based on corpus coverage and the model's Max Input, with a 30% input budget and background task progress.
• Per-library correction files, included in backups and cleaned up with library deletion.
• Schema and graph statistics APIs, optional hosted text reranker model and API Key settings.
• Parallel text/image reranking with independent relevance gates, rank fusion and text fallback.
• Search skill updates for source images, targeted graph lookups, and clearer fact-quality and historical-evidence handling.

Design choices in this release

• Types and identity: Expand the upper classes to eight and express inheritance, identity scope, and class/instance distinctions in LinkML. Together with table row extraction and external identifiers, this gives entity resolution more explicit evidence.

• Sampling and capacity: Replace a manually set sample count with automatic sampling that considers corpus size, content diversity, and model capacity. A 30% input budget balances coverage with context cost.

• Time and corrections: Current and historical views distinguish successive facts. Corrections belong to a library, survive graph rebuilds, and are removed when the library is deleted, keeping revisions reusable without passing old rules to a new library.

• Text and images: Each model judges relevance on its own scale; rank fusion combines the accepted results. Bounded image candidates and parallel scoring keep the existing request budget. The skill uses source images and focused follow-up queries where they help answer the question.

Upgrade from 0.2

1. Run app/.venv/bin/kb cleanup backup and delete existing graphs in the console. Parsed files and RAG indexes remain available.
2. Update the checkout to v0.3.0 and run ./deploy.sh with your usual deployment options.
3. On Linux, restart carrel-web and carrel-search with systemctl --user restart carrel-web carrel-search. On other hosts, restart those application processes.
4. Fill in Max Input for graph models, extract labels, and rebuild each enabled graph. Existing parsing and chunks are reused.

The graph format changes in 0.3.0. API clients use uppers in place of parent_types, upper in entity results, and the eight classes agent, entity, part, property, process, event, concept, and document. Select or upload a schema version to change types and relationships. The bundled search skill includes these changes.