A self-hosted RAG chatbot platform that connects to an organization's existing MediaWiki knowledge base, so employees can ask questions in plain language and get answers pulled directly from it in seconds - with every answer traceable to its source.
This case study documents Unitech DACH's RAG Chatbot platform, built for organizations that run substantial internal documentation on MediaWiki - enterprises, educational institutions, internal wiki teams, and legal/compliance departments that depend on a wiki-based knowledge base for day-to-day work.
For these organizations, MediaWiki serves as the system of record for policies, procedures, and institutional knowledge - but a wiki is only as useful as an employee's ability to quickly find the right page and confirm it actually answers their question.
On a traditional MediaWiki setup, employees have to browse or search for a page and then read through it - often several related pages - to confirm they've found the right answer. Search is limited to exact keyword matching, so results depend on guessing the wiki's exact wording rather than understanding what the question actually means.
Content only reflects reality when someone manually updates and re-indexes it, so a wiki can quietly fall out of date. And because access to a wiki is typically all-or-nothing, opening it up as a self-service knowledge tool means either restricting who can use it at all, or building a role- and page-level permission system from scratch - work that can take months on its own.
|
PROBLEM Employees can only search a wiki by exact keyword match, must read full pages to confirm an answer, and content drifts out of date between manual re-indexes. |
CONSEQUENCE Finding a specific answer takes browsing and reading, not asking - and building proper role-based access control for self-service search can take months on its own. |
Unitech DACH's RAG Chatbot platform turns an existing MediaWiki instance into a conversational AI assistant, without requiring the organization to migrate its documentation anywhere else.
| BEFORE | WITH THE RAG CHATBOT | |
|---|---|---|
| Finding an answer | Browse the wiki, search by exact keywords, read through multiple pages | Ask a question in plain language, get an answer in seconds |
| Search matching | Exact keyword match only | Semantic search that understands the meaning of the question |
| Keeping content current | Manual updates and re-indexing | Automatic crawling with incremental sync |
| Access control | Users must know to avoid restricted content themselves | AI automatically enforces role- and page-level permissions |
| Deployment time | Months to build a comparable system | Days to connect an existing MediaWiki instance |
The platform runs behind a FastAPI backend and processes an organization's MediaWiki instance through a five-stage pipeline: collect (automatically crawl the wiki), understand (extract and standardize content), summarize (AI-generated summaries keep the index efficient), analyze (split content into semantic segments rather than arbitrary chunks), and smart storage (index everything into OpenSearch for combined vector and keyword search).
On the query side, incoming questions are handled by a search layer that runs semantic, contextual, and keyword search in parallel, fuses the results, and reranks the top candidates with a CrossEncoder model before an answer is generated and streamed back with source citations. Role-based access control is enforced at the same stage, filtering results to what the requesting user is authorized to see before an answer is ever generated. The system runs on FastAPI, OpenSearch, and PostgreSQL behind an OpenAI-compatible streaming API, so it can be deployed self-hosted rather than routed through a third-party service.
|
<5 sec First-token response time |
>98% Accuracy with hybrid search + reranking |
100+ Concurrent users supported |
Days To deploy vs. months to build in-house |
According to the platform's stated performance benchmarks, the chatbot returns a first token in under 5 seconds and streams the rest of the answer in real time, with accuracy above 98% from combining hybrid search with reranking rather than relying on a single search method. The architecture is built to support 100+ concurrent users on an async backend, with knowledge capacity extending to hundreds of thousands of pages or documents and incremental synchronization keeping the index current without a full manual re-index. Deployment time drops from months, if an organization tried to build a comparable system itself, to days to connect an existing MediaWiki instance.
|
DEPLOYMENT OPTION Also available as a ready-to-use virtual appliance — a pre-configured virtual disk image with every required component already installed and set up — for organizations that want a fast, self-contained rollout with minimal in-house setup work. A key differentiator for customers who value deployment simplicity. |
INDICATIVE INVESTMENT Typically a fraction of the cost and time required to build and maintain a comparable system in-house. |
FastAPI (Python), PostgreSQL, OpenSearch (vector + BM25), FAISS (HNSW, 1024-dim embeddings), CrossEncoder reranker, vLLM, llama-index / sentence-transformers, mwparserfromhell, Docker Compose, OpenAI-compatible streaming API.