Glue Plus Bedrock: Offline Doc Enrichment Before Online Code RAG

Interactive code RAG should stay thin: retrieve, cite, answer. Enrichment — summaries, tags, API blurbs — belongs in an offline pipeline. Amazon Glue plus Bedrock batch (or modest on-demand) lets you improve documentation quality every night without taxing IDE latency budgets.

⚡ TL;DR: Extract symbols + raw docs to S3; Glue jobs call Bedrock to emit summaries, tags, and “when to use” notes; write enriched artifacts beside source checksums; online KB indexes enriched blobs only. Idempotent by content hash. Pair with Codebase Embeddings Refresh and Bedrock Batch Inference.

Pipeline shape

git merge → S3 raw/ → Glue enrich (Bedrock) → S3 enriched/ → KB sync → OpenSearch/pgvector

# glue_enrich.py — PySpark / Glue job sketch
import hashlib, json, boto3

bedrock = boto3.client("bedrock-runtime")
MODEL = "anthropic.claude-3-haiku-20240307-v1:0"

PROMPT = """Summarize this code symbol for developer RAG.
Return JSON: {summary, tags[], when_to_use, pitfalls[]}
CODE:
{code}
"""

def enrich_row(path: str, code: str, sha: str) -> dict:
    content_hash = hashlib.sha256(code.encode()).hexdigest()
    # ✅ Skip if enriched already for this hash
    if already_enriched(path, content_hash):
        return load_enriched(path, content_hash)
    body = {
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 512,
        "messages": [{"role": "user", "content": PROMPT.format(code=code[:12000])}],
    }
    resp = bedrock.invoke_model(modelId=MODEL, body=json.dumps(body))
    out = json.loads(resp["body"].read())
    text = out["content"][0]["text"]
    meta = json.loads(extract_json(text))
    meta.update({"path": path, "git_sha": sha, "content_hash": content_hash})
    return meta

Keep online RAG lean

Online index fields: summary, tags, symbol, path, service — not the full Bedrock transcript. Full source remains available via a tool fetch (see Tool-Augmented RAG when published, or fetch by path today).

{
  "path": "src/billing/InvoiceService.ts",
  "symbol": "InvoiceService.create",
  "summary": "Creates a draft invoice and enqueues ledger posting.",
  "tags": ["billing", "ledger", "idempotent"],
  "when_to_use": "New subscription invoice after checkout success",
  "pitfalls": ["Must pass idempotencyKey", "Do not call from sync webhooks"],
  "content_hash": "a1b2..."
}

Cost and quality controls

❌ Enriching every minified vendor file. Allowlist src/** and packages/*/src/**.

Wire to Knowledge Base sync

After enriched/ lands, trigger the same webhook/checksum sync you use for source — see sibling topic on KB git webhooks. Online Retrieve should filter on tags and service metadata for precision.

# illustrative Glue trigger after merge batch
aws glue start-job-run --job-name code-doc-enrich \
  --arguments '{"--git_sha":"'"$SHA"'","--prefix":"raw/'"$SHA"'/"}'

Closing checklist

✅ Dos
– ✅ Enrich offline; retrieve online
– ✅ Idempotent by content hash
– ✅ Validate JSON schemas before index
– ✅ Scrub secrets/PII before Bedrock
– ✅ Track token spend per repo per night

❌ Don’ts
– ❌ Don’t call Bedrock enrichment inside the IDE request path
– ❌ Don’t index raw model chatter or chain-of-thought
– ❌ Don’t enrich generated/ or node_modules
– ❌ Don’t skip checksum tombstones on deleted files

Related reading

• Codebase Embeddings: Refresh Pipelines When Git SHAs Keep Moving
• Bedrock Batch Inference: Overnight Refactor Diffs Across Large Repos
• Bedrock Prompt Caching and Batch Inference
• Secret-Aware Context Filters

Last updated on September 11, 2026