WeKnora

I used WeKnora v0.9.0 as the knowledge step of a “questions over my own documents” assistant, fed by Unlimited OCR. The Docker Compose stack started fine, but the first image pull took 45 minutes, and the UI wasn’t usable until I had fixed four things. Every agent needs its own chat model; the workspace default isn’t enough. The agent modes need a rerank model, which OpenAI doesn’t offer, so I ran a local one. WeKnora’s SSRF guard blocked that local reranker until I allowlisted it. And keyword search fails in v0.9.0 (Tencent/WeKnora#4899) unless question generation is turned on. I found most of this in the frontend source, not the docs.

Once configured, it works well. Uploads were processed in under 20 seconds. Quick Answer returned correct totals with cited chunks in about 4 seconds, and reranked search took about a second. The agent mode was just as fast, but it trusted bad OCR input: it said an item cost 1 AED because the OCR had dropped the price.

On enterprise fit: it is self-hosted, with user accounts, tenants and an SSRF guard. The GitHub license shows as NOASSERTION, so legal review would come first.