WeKnora: grounded answers, scoped memory and deliberate operations
WeKnora explained: document answers, agents, Wiki and personal memory
Understand four different knowledge behaviors and why a self-hosted RAG application is not automatically private or correct.
What you will learn
- Four jobs behind one knowledge interface
- A citation starts verification rather than ending it
- Self-hosting is a deployment choice, not an egress guarantee
Before you start
- Basic HTTP and container concepts
- Understanding of documents, passages and model providers
Separate ingestion, retrieval, answer support and memory scope, then design an evidence-based acceptance exercise.
Key takeaways
- Shared documents and personal memory have different roles.
- A citation needs claim-level verification.
- Private hosting does not automatically disable external providers.
Four jobs behind one knowledge interface
WeKnora turns uploaded or synchronized material into a searchable knowledge base. At commit 17f8938, the README describes Quick Q&A for retrieval-grounded answers, a ReAct agent for multi-step tool use, Wiki Mode for maintained Markdown knowledge, and cross-session memory for personal context. These are related capabilities, not four names for the same search.
Consider a synthetic product handbook: Quick Q&A locates a documented upload limit, an agent compares several sources, Wiki organizes recurring concepts, and memory remembers a reader’s preference for concise explanations. The handbook is shared evidence; the personal preference is not a new fact about the product.
A citation starts verification rather than ending it
An answer with a clickable source is easier to inspect, but the source must actually support the associated claim. Open the cited passage, check its version and distinguish quoted policy from the model’s inference. A relevant document can still be attached to an unsupported conclusion.
The quickstart separates a conversation model from an embedding model, with rerank and multimodal components optional. The application is not a bundled guarantee of model quality. Document parsing, candidate retrieval and answer composition can fail independently, so diagnose the stage before changing the model.
Self-hosting is a deployment choice, not an egress guarantee
The framework supports private deployment and replaceable providers. A remote embedding API, external chat model, web search service or remote sandbox can still receive data. Draw the actual data path before promising that documents never leave the machine.
This series uses fixed repository documents and source inspection plus twelve executed tests of a synthetic recall model. We did not install WeKnora, upload documents, call models, start a sandbox or benchmark a database. Source-backed behavior and proposed exercises remain clearly separated.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Choose a concrete knowledge question.
- 2
Decide whether retrieval, tools or curation are needed.
- 3
Map source ownership and outbound model calls.
Copy-ready example
{
"syntheticHandbook": true,
"sharedEvidence": "upload policy",
"personalPreference": "concise explanations",
"citationChecked": false,
"remoteProvidersReviewed": false
}Frequently asked questions
Is Wiki Mode just another vector index?
No. The README describes generated, editable Markdown pages with revision history.
Does self-hosting mean no external data transfer?
Only if the configured models, integrations and runtime also stay inside the intended boundary.
Sources
- WeKnora / README.mdSource checked 2026-09-14
- WeKnora / website-docs/01-getting-started/03-quickstart.mdSource checked 2026-09-14
- WeKnora / website-docs/03-features/23-memory.mdSource checked 2026-09-14