Aviation knowledge system

A technical-document retrieval workflow with answers linked to sources and relevant video.

I designed the retrieval workflows, prepared document routes and built SharePoint synchronization with the technical team.

AI Engineer Intern

Garuda Indonesia

Nov 2025 to May 2026

Project completed Jun 2026

System overview

Input: SharePoint documents, technical questions and optional uploaded documents
Processing: File-specific preparation, retrieval and Qwen answer generation
Output: Answers with source references and relevant video where supported

Answer examples

I designed the retrieval and response workflows and built SharePoint synchronization; teammates provided the Dify infrastructure and chat website.

Team-built chat screen with a technical answer and a linked PDF source. — A chat example connects the answer to a technical PDF.

Team-built chat screen showing an answer with a related training-video preview. — A related video appears with the response.

Technical answers require review against the cited source.

System

SharePoint synchronization prepares incoming files for the knowledge base. At question time, the workflow retrieves context from that index or from a document uploaded with the question.

Architecture diagram connecting SharePoint preparation, vector knowledge storage, Dify retrieval and a cited response. — The architecture separates keeping the knowledge base current from retrieving context for a question.

Redacted Dify workflow showing parallel document-extraction routes joining before chunking and knowledge-base indexing. — Redacted workflow view showing parallel extraction paths joining before indexing.

Path	Engineering work
Ingestion	Synchronize SharePoint files; prepare formats separately before chunking, embedding and indexing.
Question routing	Distinguish standing-knowledge queries from questions with an uploaded document.
Retrieval and generation	Retrieve relevant passages and tune Qwen parameters for technical responses.
Response assembly	Combine answer text with source references and optional video output for the chat interface.

Workflow implementation

Redacted Dify canvas showing the workflow from input routing through retrieval and response generation. — The redacted canvas shows the retrieval and response path in one workflow view.

Redacted Dify branch distinguishing a knowledge-base query from a question with an uploaded document. — The route separates indexed knowledge search from an uploaded document.

Redacted Dify branch checking whether to retrieve a relevant training video. — A conditional branch requests a video when the response route needs one.

Redacted Dify output node combining answer text with source and optional video fields. — The output path combines answer text with source references and optional video data.

Development trial

Trial scope	Development trial outcome
About 100 documents, roughly 100 to 500 pages each, with 20 prepared questions per document.	18 to 20 answers judged correct out of 20 for a tested document after adjustments. Author-reported; no held-out corpus or documented scoring rubric. This does not establish general retrieval accuracy or aviation safety.