Glider Docs Copilot

Glider Docs Copilot is a local-first documentation assistant for the Glider IDE GitBook.

Instead of wiring a hosted chatbot to docs, this project builds a searchable local knowledge base from the Glider documentation and lets users ask natural-language questions against it through a simple chat interface.

What It Does

• Scrapes and normalizes the glider-ide GitBook into a local dataset.
• Splits documentation into semantic chunks for better retrieval.
• Generates local vector embeddings for every chunk.
• Embeds the user query in the browser.
• Returns the most relevant documentation sections with source links.
• Runs without a server-side LLM answer pipeline.

The result is closer to “talk to docs with grounded retrieval” than a generic AI chatbot.

Why Xenova Matters Here

This project relies on the Xenova/Hugging Face browser model:

• Model: Xenova/all-MiniLM-L6-v2
• Runtime package: @browser-ai/transformers-js

That model powers the semantic search layer in two places:

1. Offline indexing

   • The project embeds every Glider docs chunk ahead of time and stores the vectors in data/glider-vectors.json.

2. Runtime query understanding

   • When a user asks a question, the app generates a local embedding for the query in the browser and compares it against the stored vectors.

Why this is useful

• No OpenAI or hosted embedding API is required for search.
• First-party docs stay local to the app flow.
• Search works by meaning, not only exact keyword matches.
• The model is cached after first load, so repeated use is smoother.
• It is especially good at mapping natural questions to related technical documentation sections.

What Xenova is not doing here

• It is not generating long-form answers with a remote LLM.
• It is not calling a cloud inference endpoint.
• It is not replacing the docs source of truth.

Its role is retrieval: turning both docs and user questions into vectors so the app can find the right documentation quickly and locally.

Architecture

Data pipeline

• scripts/scrape-glider-docs.ts
  • pulls Glider GitBook markdown from the sitemap
  • normalizes content into pages and chunks
• scripts/generate-embeddings.ts
  • generates embeddings for every chunk using Xenova

App runtime

• components/GliderDocsCopilot.tsx
  • chat UI, suggestions, source cards, follow-ups
• lib/semantic-search.ts
  • local embedding + cosine similarity search
• lib/prompts.ts
  • lightweight answer/follow-up shaping

Local Development

Install dependencies:

npm install

Start the app:

npm run dev

Open:

http://localhost:3000

Rebuilding the Docs Index

If the Glider GitBook changes, rebuild the local corpus:

npm run build-data

That runs:

npm run scrape-docs
npm run generate-embeddings

Production

This project is deployed on Vercel:

• Live app: https://glider-docs-copilot.vercel.app

Current Tradeoffs

• Retrieval is semantic-only right now; hybrid keyword + semantic search would improve exact API lookups.
• Answer synthesis is intentionally lightweight and grounded in retrieved chunks.
• The first load can be slower because the local embedding model needs to warm up and cache.

Good Next Steps

• Add hybrid retrieval for exact method names and symbols.
• Add reranking on top of the Xenova recall layer.
• Add metadata-aware boosting for api/ pages versus tutorial pages.
• Add tighter chunking for code-heavy API entries.