Project three a Workers AI and Vectorize knowledge assistant

Build minimum RAG with multilingual embeddings, traceable metadata, protected ingestion, and an explicit no-context branch.

PROJECT 03ADVANCEDAbout 70 minutesOutcome: minimal sourced RAG

Done means

/index and /ask require separate secrets and Vectorize metadata traces to source text; queries pass only embedding.data[0], while no matches return 404 no_context instead of inviting unsupported generation.

Workers AI and Vectorize knowledge assistant architecture — Ingestion generates vectors for asynchronous upsert; query retrieves metadata before sending constrained context to generation.Open original

01
Ingest
A secret-protected document endpoint.
02
Embed
bge-m3 produces multilingual embeddings.
03
Retrieve
Vectorize returns source-text metadata.
04
Answer
Generation runs only with matched context.

Files and resources

README.md

wrangler.jsonc

index.mjs

Runnable source lives in examples/ai-knowledge-assistant/. INGEST_KEY and QUERY_KEY are secrets and must not be vars values.

Prepare and verify

Create the multilingual vector index

pnpm wrangler vectorize create orange-book-knowledge --preset @cf/baai/bge-m3

This creates a remote resource and requires account authorization; confirm the target account first. --preset lets Wrangler derive dimensions and metric from the current embedding model.

Set the secret and start

cp examples/ai-knowledge-assistant/.dev.vars.example examples/ai-knowledge-assistant/.dev.vars
pnpm wrangler dev --config examples/ai-knowledge-assistant/wrangler.jsonc

Replace both placeholders in .dev.vars and never commit that file. Workers AI and Vectorize use remote service quota and are not a fully offline preview; set production secrets separately only after confirming the target account.

Index one document

curl -i http://127.0.0.1:8787/index \
  -H 'content-type: application/json' \
  -H 'x-ingest-key: <your-local-secret>' \
  --data '{"documents":[{"id":"tls-strict","text":"Prefer Full strict for a new site.","source":"/en/docs/fundamentals/tls-modes"}]}'

After 202 indexing, wait a few seconds before querying because Vectorize upsert becomes visible asynchronously.

Ask and inspect sources

curl -i http://127.0.0.1:8787/ask \
  -H 'content-type: application/json' \
  -H 'x-query-key: <your-local-query-secret>' \
  --data '{"question":"Which TLS mode should a new site use?"}'

Confirm sources came from indexed metadata. With no indexed match, expect no_context, not a fabricated answer.

Rollback and security boundary

Remove public routing or roll back the Worker version first; back up or rebuild index data separately.
Rotate a leaked INGEST_KEY or QUERY_KEY; deleting code is insufficient.
QUERY_KEY is only a server-to-server tutorial guard and must not be embedded in a browser; a public UI also needs user authentication, abuse protection, and cost limits.
Model names are time-sensitive configuration. Recheck the model catalog, pricing, and data policy before release.
Retrieved content enters the prompt as untrusted reference data. This example returns text and never executes model- or document-proposed actions.