RAG: the pages it generated
Retrieval-augmented generation over this dealer's data. Every vehicle and store page is expanded into plain-text Markdown you can browse below, pull as one bulk corpus, or query live with the rag_search tool.
How the RAG corpus works
Each vehicle and store page is written out as clean Markdown and concatenated into a single bulk corpus, llms-full.txt (530 KB). An agent can ingest the whole corpus, or call rag_search(query) over the MCP endpoint to retrieve just the matching chunks with source links. Everything below is a real generated page you can open.
Every document is checked against an admission policy before it joins the corpus. Search-results and filter pages are refused, interface furniture and text that repeats across the corpus is removed, thin and near-identical documents are refused, and each refusal is recorded with its reason.
| Document type | Documents |
|---|---|
| site | 17 |
| Measure | Value |
|---|---|
| Documents evaluated for admission | 61 |
| Admitted | 17 |
| Refused | 44 |
| Average unique words per admitted document | 188.6 |
| Repeated blocks lifted into a single document | 43 |
| Near-duplicate ratio | 0.033 |
| Reason | Documents |
|---|---|
| Too little unique content | 33 |
| Mostly repeated boilerplate | 8 |
| near duplicate of finance-locations.md jaccard | 1 |
| Navigation or utility page | 1 |
| near duplicate of scheduleservice.md jaccard | 1 |
All 474 vehicles are in the corpus as one Markdown record per VIN, addressed by the pattern https://ai.castlecars.com/v/{VIN}.md (and https://ai.castlecars.com/v/{VIN}.json). A few samples:
rag_search(query, limit?) over the MCP endpointhttps://ai.castlecars.com/v/{VIN}.mdcurl -s -X POST https://ai.castlecars.com/api/ucp/mcp \
-H 'content-type: application/json' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"rag_search","arguments":{}}}'