Disclaimer, company use. Xia Wu is a RAG agent for internal use, a platform built for company use. This page exists to document the work and its author: it presents only the structure of the system and a plain-language account of what each part does. No source code, prompt text, configuration values, credentials, or business data are published here or anywhere on this site. A small number of file names have been genericised where the original named a third-party service or a customer sector.
In one paragraph
Xia Wu is a RAG agent for internal use, an AI work platform for pre-sales teams. Retrieval augmented generation is the whole design: it answers from the company’s own documents rather than from the model’s memory, and shows where every answer came from. It turns a company’s scattered documents (Word files, PDFs, slide decks, spreadsheets, images) into a searchable knowledge base that answers questions with verified citations, and then uses that same knowledge to carry a project through the whole pre-sales document chain: requirement research, solution architecture, and full solution writing. Finished official documents flow back into the knowledge base automatically, so the platform learns from its own output.
By the numbers
234Files
85,355Lines
298Relationships
28Test suites
Built and hardened across nine documented delivery phases, from a single-machine Windows prototype to a server-ready platform designed for around one hundred accounts with fifty concurrent users.
Core capabilities
Knowledge & answers
Ingests Word, PDF, PowerPoint, Excel and image sources while preserving page and slide positions. Retrieval runs keyword and vector search together over a dedicated vector database, and every answer is checked so that each citation genuinely points at retrieved text. The system is built to refuse to invent its sources.
Vision auto-chunking
An agent that renders document pages and slides as images, asks a vision model to describe what they show, and proposes richer knowledge entries for a human to approve. It replaced a fully manual annotation workflow, and runs as a resumable job with a review checkpoint in the middle.
Governed intake
Any account can upload material, but nothing enters the knowledge base without an administrator’s approval. A sensitive-word screen covers user input, model output, web search results and generated documents, and every hit is logged.
Project workflow
Projects are the spine of the platform: created under immutable names, typed from a maintained list, each with its own folder structure, activity record and history knowledge area. Every downstream document belongs to a project.
Requirement research
Generates research templates, absorbs uploaded source material, and transcribes recorded client meetings through a cloud speech service, then produces a structured research report that can be refined section by section.
Architecture design
Selects the right architecture template for the project’s type, drafts the solution architecture, and supports refinement rounds and manual edits before the approved version is archived.
Solution writing
The largest engine in the platform: builds a complete solution document chapter by chapter, analyses and imports existing drafts, generates guidance for the figures each section needs, drafts slide outlines, and renders the finished file for download.
The learning loop
Each project function keeps exactly one active official version. Once an official document has been stable for its holding period, it is pushed automatically into the knowledge review queue. Today’s deliverable becomes tomorrow’s reference material.
Task platform
Long jobs run on a background worker platform with queues, document locks, live progress streamed to the browser, idempotent starts (a double-click never doubles the work), cancellation, retry, and human review steps. Runs checkpoint themselves so they can resume instead of restarting.
Identity & security
Role-based accounts and permissions, hashed passwords, session and anti-forgery protection, security headers, an append-only audit trail, and each user’s own model credentials held encrypted at rest under a portable master key.
Integrations
A surface for other systems: service-account tokens stored only as hashes, registered webhooks with signed payloads, and a delivery outbox that retries until events land.
Proof, not vibes
Twenty-eight regression suites, frozen retrieval baselines, smoke checks over the key pages, and a baseline verifier that detects drift from the signed-off system state.
Operations
One-click launchers, health probes, schema migrations, and a documented path from the original Windows prototype to Linux server deployment on PostgreSQL, Redis, a vector store and object storage.
How to read the map
The layout is not decorative. It was produced by a force simulation over the project’s real dependency graph, extracted by parsing every import, template load, launcher script and test in the build. Files that depend on each other pull together; the nine subsystems settle into their own regions of sky. The hub stars with the most filaments are the load-bearing pieces: the shared error contract, the runtime configuration, the model gateway. Hover any star to isolate its connections; select one to list every file it depends on and every file that depends on it.