The RAGFlow v1.0.0 stable release helps teams building answers from retrieved documents keep retrieval within users' access permissions. It also fixes document structure preservation and agent execution errors, addressing incorrectly split documents and empty execution results in retrieval-augmented answering.

1. Apply permissions and caps to retrieval scope

Retrieval now caps candidate counts for reranking results and nearest-neighbor searches for similar documents. This limits the candidate pool reviewed before documents are selected for the final answer.

SQL queries in conversational retrieval now include filters for knowledge bases the user can access. A knowledge base is a space for storing documents. Permission filters protect the scope of searchable documents, while candidate caps limit the size of the pool under review. Deleting document fragments, or chunks, is also restricted to the specified dataset, which groups documents for management. Updating a single chunk preserves its dataset and document scope.

The Dify integration was fixed so that empty filters on metadata, the descriptive information attached to documents, do not narrow retrieval scope.

The four tools named in the release, including arxiv and duckduckgo, have a response body cap of 4 MiB. This limits the size of the response received from a tool.

List APIs have a page size cap of 100. This applies to the number of items fetched at once. The response body cap measures bytes, while the page size cap counts items, so the two figures cannot be compared directly.

2. Fix document parsing and splitting errors

Sentences cut off at PDF page boundaries were fixed, and missing table rows and merged headers are now preserved. Fixes also address PDF child chunks overwriting one another and source locations not matching previews, reducing errors in document structure handling before retrieval.

Document processing improvements span parser options, ingestion paths, and external API compatibility. MiniCPM-OCR was added as a PDF parser, a program that reads document structure, while MonkeyOCR supports PDF processing in the Go-based document ingestion path. The MinerU parser improves compatibility with the 4.1 job API and allows timeouts to be adjusted through environment settings.

Support was also added for canceling the entire document processing queue and canceling running jobs. This reduces the burden of canceling queued documents one by one when incorrect jobs accumulate.

3. Improve agent execution and external connections

Agents now correctly locate components of nested sub-agents.

A bug that left structured output, which returns results in predefined fields, empty was also fixed.

Crashes at execution state checkpoints were prevented, and component and conditional branch errors now propagate to the caller. This helps trace which stage failed in workflows that connect multiple steps.

Translation and weather credentials now come from node settings or dedicated environment variables, separate from model-visible parameters. This keeps model inputs and external service credentials managed separately.

Refining retrieval permissions and input caps makes it easier to manage retrieval and tool call scope. Teams whose setups assume large candidate pools or responses will need to review their existing settings and request sizes.

4. Install the command-line tool for your environment

The stable release assets include 6 command-line tool builds for different operating systems and processors, installation scripts, and a checksum file.

The shell installer detects the operating system and architecture, downloads the matching file, and checks its checksum, a digest used to verify file integrity, against SHA256SUMS before installation. The default installation directory is /usr/local/bin, and the settings below are taken directly from the script.

INSTALL_DIR="${INSTALL_DIR:-/usr/local/bin}"
GITHUB_REPO="${GITHUB_REPO:-infiniflow/ragflow}"
CLI_NAME="${CLI_NAME:-ragflow-cli}"
VERSION="${VERSION:-latest}"

The default installation version points to the latest release. To pin this stable release, specify the version setting during installation; the path can also be changed as needed. A Windows installation script is also available among the release assets.

v1.0.0 focuses on making retrieval, parsing, and agent execution more predictable, so it is worth reviewing both the tighter scope limits and the installation method before upgrading.