Archat

Built around building specifications

Archat explores a specific document workflow: finding relevant material in Danish building specifications, asking questions about it and using existing material to draft a Word document. It is a small MVP using a 4B-parameter Qwen3.5 model, with the Jetson Orin Nano as its intended compact runtime.

What the implementation includes

• PDF and Word document ingestion into a searchable document collection.
• Streaming chat through a local Ollama model.
• Keyword retrieval with BM25 and optional semantic retrieval, combined into one ranking. The code falls back to keyword search when the semantic dependencies are unavailable.
• An experimental Word-template generation path that retrieves material to help fill document sections.

From retrieval to a trustworthy answer

The chat implementation passes source filenames and retrieved passages into the model context. It does not yet provide a separate, structured citation response. Useful retrieval and a source-aware prompt still need to be tested for accurate answers, inspectable references and sensible handling of missing information.

The drafting path exists in the code; that does not establish the correctness of its output or preservation of every template detail. A reviewer needs to check generated documents against the source material and the intended specification.

What has been checked

A local source review on 15 September 2026 confirmed these implementation paths. The configured Qwen3.5 4B model was unavailable in the review environment, so no end-to-end inference or Jetson latency result is claimed here.

A focused next step

Evaluate one approved document set with representative questions, including questions the documents cannot answer. Check retrieval relevance, references, template output and response time on the intended hardware before expanding the scope.

Local inference is a deployment choice. Access controls, document handling and support still need to be agreed for any client use.