carrel
Ontology-Augmented Generation for AI agents.
Automatically checked · GitHub checked 2026-11-16

Overview
Carrel
English | 中文
Ontology-Augmented Generation for AI agents.
Carrel is a self-hosted knowledge base for AI agents. It automates document parsing, indexing, and retrieval, with optional knowledge graphs. Agents use its API to retrieve source passages, images, and related facts, then generate their own answers.
Use it to give an agent access to a collection of product documents, research, contracts, meeting notes, or other files that change over time. A web console manages knowledge bases, models, and processing tasks. A bundled search skill connects agents such as Claude Code and Codex to the retrieval service.
0.4.1 general retrieval benchmarks and codebase capabilities · Full 0.4.0 capability evaluation
Features · Evaluation · Design · Quick start · Agent integration · Documentation
Features
0.4.3 changes and upgrade steps — Consistent document publication, version-bound evidence, document and section time context, table parsing repairs, and bundled code grammars.
- Document parsing: PDF, Word, PowerPoint, spreadsheets, images, HTML, Markdown, notebooks, source code, and workflow files, with processing suited to each format. Format details.
- Automatic updates: scan for file changes, update indexes, and maintain enabled knowledge graphs through scheduled tasks.
- Hybrid retrieval: combine vector, keyword, graph, and visual search, with automatic knowledge-base selection and optional text/image reranking with rank fusion.
- Knowledge graphs: extract entity types, relationships, and facts from your documents; retain sources, time information, and conflict indicators, compare graph versions, and apply reviewed, evidence-scoped corrections.
- Agent access: search through an API, fetch surrounding text or original images, and follow graph relationships when more evidence is needed.
- Web console: configure processing, inspect chunks and graphs, track tasks, and manage services. Available in English and Chinese.
Capability evaluation
Across 275 text queries, 0.4.1 reached 93.45% evidence Hit@5 and 0.8141 MRR@10. In a separate study of 24 code questions, Haiku answered 13 completely from the first response and 23 with Skill follow-up. 0.4.1 evaluation
The full 0.4.0 evaluation also covers graph rebuild stability, time-scoped evidence, conflicts and ontology governance. Full 0.4.0 technical report
How it works
flowchart LR
D[Document folders] --> P[Parse and chunk]
P --> I[Vector and keyword indexes]
P --> G[Optional knowledge graphs]
I --> R[Retrieval API]
G --> R
R --> A[AI agent]
A -->|Follow-up queries| R
Carrel uses MinerU and native parsers for ingestion. Standard uses Qdrant, OpenSearch and Neo4j; Light uses embedded LanceDB and DuckDB. SQLite tracks files and tasks in both editions. Model services run independently, locally or through compatible hosted endpoints. See Architecture for the pipeline and retrieval design.
Design
Carrel provides general knowledge storage, retrieval, and evidence interfaces. Entities, relationships, and facts retain sources, time information, and conflicts so agents can check and use them alongside the original documents. Default retrieval retains historical and future source claims. An explicit time query selects the relevant evidence; reviewed corrections record which sources and identities the decision applies to. The Agent Loop plugin is in development and not yet implemented. It is intended to use these interfaces to continuously learn an enterprise ontology, seek clarification from relevant people, and maintain correction rules.
Routine graph updates merge new and changed content incrementally. Configured time or change thresholds trigger full rebuilds, balancing update cost with overall consistency. Parsing, indexing, and graph construction are maintained separately: rebuilding a graph can reuse parsed content and chunks while preserving RAG indexes.
Requirements
- Docker with Compose v2. Python and background scheduling run inside the application container.
- A text embedding endpoint; a vision-language endpoint for image descriptions; and a chat model for knowledge graph construction, if enabled.
- An NVIDIA GPU and NVIDIA Container Toolkit for GPU parsing. CPU parsing is also supported. Resource needs depend on the parser and collection size.
| Platform | Deployment |
|---|---|
| Linux with NVIDIA GPU | GPU parsing with NVIDIA Container Toolkit; container scheduling |
| Linux without GPU | CPU parsing; connect to model services on another host or a provider |
| macOS with Docker Desktop | CPU parsing and application containers, including scheduling |
| Windows | Use WSL2; native Windows deployment is not supported |
Example setup describes the author's installation on an NVIDIA DGX Spark and some lighter alternatives.
Quick start
1. Deploy the services
git clone https://github.com/FengZhenfei/carrel.git
cd carrel
./deploy.sh --standard --build
# Or use embedded storage:
# ./deploy.sh --light --build
The optional script prepares Compose configuration and starts the application, background scheduler and selected dependencies. You can also use Compose directly. First startup downloads MinerU parser assets. See Container deployment.
2. Configure model endpoints
Choose Standard for Qdrant, OpenSearch and Neo4j, or Light for embedded LanceDB and DuckDB. The edition is fixed for that deployment directory. Both include MinerU; Light has no database containers. See Storage editions.
In System settings, configure LLM, Embedding and Reranker endpoints, model IDs and API keys. Register a vision-capable LLM and select it as the Multimodal model for each knowledge base. Images, charts and table review share that selection. Models can run on your own machines or hosted platforms. Carrel does not start these model services or download their weights. See Model configuration.
3. Add documents
Open the console at http://127.0.0.1:9800. Put documents in
data/runtime/mirror/<folder>/, then enable that folder in the console. Each
enabled folder becomes a knowledge base; scheduled tasks process its files.
For a knowledge graph, register a chat model with its Max Input in the console, select it in the knowledge-base settings, and enable graph building. Check service health and task progress in the console before making your first query.
Agent integration
Install the bundled carrel-search skill in your agent, or call the API directly:
The first container start creates KB_SEARCH_TOKEN in
data/config/knowledge-base.env. Set CARREL_SEARCH_TOKEN to that value:
export CARREL_SEARCH_TOKEN='your search token'
curl -sS http://127.0.0.1:9810/search \
-H "Authorization: Bearer $CARREL_SEARCH_TOKEN" \
-H 'Content-Type: application/json' \
-d '{"question": "What are the delivery and acceptance requirements?", "top_k": 8}'
For a remote agent,
configure CARREL_SEARCH_BASE_URL and CARREL_SEARCH_TOKEN in its environment.
Responses include source passages and, when available, graph evidence and image information. Agents can request more context, inspect images, and follow relationships before answering. See Retrieval and API for authentication, response status, and citations.
Documentation
| Guide | Contents |
|---|---|
| Configuration | Model endpoints, folders, per-base settings, and access control |
| Architecture | Parsing, graph construction, retrieval, and code layout |
| Retrieval and API | Agent setup, endpoints, evidence status, and evaluation |
| Retrieval and codebase evaluation (0.4.1) | Document retrieval, page-image location, Skill follow-up and cross-file code analysis |
| Capability evaluation (0.4.0) | Full technical report: retrieval, answers, temporal evidence, conflicts, ontology, rebuild stability, latency, and token usage |
| Operations and development | Console, CLI, scheduling, troubleshooting, and tests |
| Test scope | Shared regression contracts, deployment-specific checks, and live evaluations |
| Example setup | The author's installation on a DGX Spark, a Mac that syncs the documents, and lighter alternatives |
| Docker Compose | Infrastructure, model weights, and container settings |
| Container deployment | Images, scheduling, persistent data and updates |
| systemd | Native/source installation |
Data and access
The console binds to localhost by default. For LAN access, set both
KB_WEB_HOST and KB_WEB_TOKEN. Configure KB_SEARCH_TOKEN before using
protected search endpoints remotely. The stores and model servers listen on
loopback only, which keeps other machines out but not web pages opened in a
browser on the same host, so avoid browsing the web there.
Documents and images are sent to the model endpoints you configure. Local endpoints keep model processing within your own infrastructure. See Access and data handling.
License
Carrel is licensed under MIT. Dependencies and model weights have
their own licenses; see Third-party notices. Original-resolution
PDF image retrieval uses an optional dependency installed with
./deploy.sh --with-pdf-images; without it, retrieval uses cached images.