More than chat with your files
WeKnora starts where basic document chat tools usually stop. Uploads are parsed, split into searchable chunks, indexed, and exposed through question answering with citations. On top of that base, Tencent has added a ReAct-style agent that can combine retrieval with web search and MCP tools, plus a wiki mode that turns source material into linked pages. Editors can revise generated pages and individual retrieval chunks, compare versions, and roll changes back.
That breadth is the main reason to consider it. A company could use one installation for an internal policy assistant, a public support widget, an agent connected to team chat, and a maintained project wiki. Multiple workspaces, four user roles, per-resource ownership, audit logs, scoped API keys, and external identity support make the pitch credible for organizations rather than only solo experiments. The API, command-line client, Chrome extension, WeChat client, and official MCP package provide several ways to get information in and out.
The tradeoff is easy to miss in the screenshots: WeKnora is a platform. It owns document conversion, background queues, model configuration, retrieval, storage, chat, authorization, and integrations. Adopting it means accepting that platform boundary and operating the services behind it.
The knowledge workflow is unusually editable
Most RAG products let administrators tune chunk sizes, then hide the resulting index. WeKnora exposes more of the work. Its folder view retains uploaded directory structure, upload batches can override parsing and chunking choices, and users can edit chunks with revision history. Automatic reindexing after an edit turns correction into a normal workflow instead of a database repair. Generated wiki pages also have line-level diffs and rollback.
Those controls matter because document retrieval is never entirely automatic. A poor table extraction, misleading heading boundary, or outdated paragraph can ruin an otherwise good answer. WeKnora gives a knowledge manager somewhere to inspect and correct that source layer. It also supports FAQ, document, and wiki knowledge bases, along with imports from services including Feishu, Notion, Yuque, and RSS. Support for common office, web, image, ebook, and structured-data formats is broad.
Breadth is not the same as dependable handling of every file. An open report against version 0.7.2 shows XLSX ingestion failing when a workbook contains charts, with the failure traced to the document reader's openpyxl save path. Another open report says AI-generated questions for one chunk can include text from neighboring chunks. Teams should build a test collection from their ugliest real files, including charts, scans, merged cells, and long tables, then inspect both parsed text and retrieved answers.
Deployment starts quickly, operations do not
The local instructions are refreshingly short: clone the repository, copy the example environment file, pull images, and start Docker Compose. The default stack includes separate frontend, application, document-reader, PostgreSQL, and Redis services. It gets a browser evaluation running without requiring Kubernetes, and the bundled database already supplies the default vector-search path.
Useful options expand the footprint. Compose profiles can add Neo4j, MinIO, SearXNG, several alternative vector databases, Langfuse and its supporting stores, and the MCP service. Local Ollama is supported, while many hosted model, embedding, reranking, web-search, object-storage, and messaging providers can be configured. This flexibility is excellent for fitting existing infrastructure, but every substitution adds credentials, network paths, compatibility decisions, and failure modes.
Production therefore needs more than copying the demo environment. Operators must pin a release instead of relying on latest, protect secrets, terminate TLS, back up persistent data, monitor queues, budget model usage, and test upgrades. The README warns that running Compose without pulling can leave cached images and the UI out of sync. The project offers Helm, private deployment, and offline support, but those are deployment routes, not substitutes for capacity and recovery planning.
The sharp edges concern data operations
The most serious current cautions are about maintaining a live corpus. One open request says a large set of failed documents cannot be retried all at once and that the UI limits each retry batch. That is manageable for a pilot and painful when a provider outage leaves thousands of jobs failed. The runtime dashboard and worker-pool controls are valuable, but bulk recovery still deserves a rehearsal.
A separate v0.7.2 report is more consequential: the reporter says deleted documents reappeared and related vector, graph, wiki, and Redis data remained until manually removed. It is one report, not proof that every backend and configuration loses deletion requests, but regulated or frequently changing corpora cannot shrug it off. Test deletion end to end against the exact storage combination you intend to use, and verify backups and indexes rather than trusting the success message.
The license also deserves a careful read. The project describes itself as MIT licensed, while its LICENSE file separately lists third-party components under their own terms. That is normal for a large distribution, but compliance teams should review the complete list and the models or services they add.
Healthy, fast-moving, and best piloted first
The repository was pushed on August 11, 2026, four days after version 0.7.2 was published. Issues and pull requests were also being opened, discussed, and merged that day. That is strong evidence of an active project. GitHub's combined open count was 540, consisting of 338 issues and 202 pull requests when checked, so it reflects both a substantial support queue and a very busy development branch.
Documentation is another strength. The English README provides a useful map, localized readmes serve three additional languages, and the project points to a much larger product documentation site covering architecture, deployment, features, APIs, clients, and extension points. The release notes are unusually specific about fixes and internal changes.
WeKnora is worth piloting when an organization wants one governed home for retrieval, agents, and editable knowledge. It offers more administrative and curation depth than a simple local chatbot. Do not mistake feature coverage for finished operational proof, though. A serious evaluation should test parsing, deletion, failed-job recovery, model spend, permissions, and upgrades with production-like data before anyone depends on it.