PRIMARY AUDIENCE
Website visitors asking about custom software engineering, architecture, cloud choices, technology orchestration, and documented Neubofy services; also visitors who need to be directed to a human support channel.
Edge-native AI · Retrieval-augmented support
NASA — Neubofy Artificial Support Agent
AI system
A technical concierge that answers from a maintained knowledge base, not a blank prompt.
Interface overview
Cloudflare Workers + Workers AIRuntime design
Cloudflare Vectorize: neubofy-kb-indexRetrieval index
One-script universal widget documentedEmbedding
01 / WHY
The problem I chose to solve
The most useful way to understand this project is to start with the friction behind it, then examine the system choices made to address that friction.
A technical support assistant needs more than a language model: it needs the organization’s specific service definitions, architecture context, support boundaries, and a safe route to a human when the question is too complex. Repeating the same foundational answers manually also scales poorly.
NASA is designed as an edge-native support layer for Neubofy. The frontend can be embedded with a single script tag; the Worker retrieves semantically relevant knowledge from a Vectorize index and streams an answer using Workers AI. A stateless runtime model avoids keeping a long-lived per-chat server object, and a documented optimization skips embedding retrieval for trivial greetings.
02 / WHO IT SERVES
Where it fits in the real world
A strong implementation has a bounded audience and context. These are the use cases supported by the current design—not a claim that all of them have been validated at scale.
WHAT MAKES IT DISTINCT
- Retrieval from the Neubofy Knowledge Base instead of relying only on model memory
- Embeddable widget for arbitrary sites without copying the widget bundle into the host project
- Cloudflare-native edge execution and Workers AI
- Designed escalation to human channels for complex requests
- Greeting shortcut intended to reduce unnecessary embedding work
03 / PRODUCT SURFACE
What the system actually does
An overview of key capabilities and system features.
CAPABILITY / 01
Universal embed
The documented snippet loads `https://nasa.neubofy.in/widget.js` and can be used with static HTML or Next.js.
CAPABILITY / 02
Semantic retrieval
Queries the `neubofy-kb-index` Vectorize index to collect relevant source context.
CAPABILITY / 03
Streaming responses
The README describes streaming responses from the Worker runtime.
CAPABILITY / 04
Support positioning
The assistant focuses on engineering, cloud architecture, playbooks, and project scoping.
CAPABILITY / 05
Search metadata
The repository describes structured organization metadata and open-graph/answer-engine markup.
04 / SYSTEM MAP
Architecture & data flow
Architecture and system flow overview.
LAYER / 01
Website visitor
Opens the standalone assistant or a site with the widget script.
LAYER / 02
Widget client
Sends the user prompt and manages the chat interface.
LAYER / 03
Cloudflare Worker
Orchestrates retrieval, prompt construction, model call, and streaming.
LAYER / 04
Vectorize + Workers AI
Retrieval supplies knowledge context; the model composes an answer that is returned to the visitor.
IMPLEMENTATION STACK
Cloudflare Workers, Workers AI, Cloudflare Vectorize, React 19, Vite, AI SDK, Cloudflare Agents packages, TypeScript, server-served widget.js.
Cloudflare WorkersWorkers AICloudflare VectorizeReact 19ViteAI SDKCloudflare Agents packagesTypeScriptserver-served widget.js.
05 / ENGINEERING JUDGMENT
Decisions behind the implementation
Engineering choices and design decisions behind the implementation.
- Ground answers in the team’s own knowledge instead of relying on generic model familiarity.
- Avoid requiring a separate integration build per customer website.
- Keep the core runtime edge-native and stateless by design.
- Use a human escalation route where a model answer alone is not appropriate.
06 / SCOPE & BOUNDARIES
Technical Considerations
Key technical details and system scope considerations.
- A live domain and repository documentation are evidence of deployment intent, not proof of accuracy, uptime, or successful handling of every query. The automated browser could not inspect the rendered UI, so a manual production smoke test is recommended.
- “Stateless” does not automatically mean “no user data is processed”; requests still pass through the model and retrieval provider. Publish retention and provider-processing details that match deployed settings.
- RAG improves relevance but does not guarantee factual answers; users should have a visible path to source documents or human follow-up.
07 / NEXT ITERATION
What I would improve next
The next step is not simply adding features. It is increasing reliability, evidence quality, user trust, and clarity about the boundaries of the system.
NEXT / 01
Add user-visible citations/source snippets to knowledge-grounded answers.
NEXT / 02
Run adversarial retrieval tests for stale articles, prompt injection in source content, and unsupported claims.
NEXT / 03
Publish a clear retention statement and monitor response quality without collecting unnecessary conversation content.
08 / SOURCES
Sources & Resources
Project repositories and resource links.