[Bug]: agentic_rag template tries to deploy an empty Vector Search index, which Vertex AI doesn't allow

[luke-swarm-hub]

File Name

not a notebook issue

What happened?

I’ve encountered an issue with the starter kit template for template 3 agentic_rag. The template tries to deploy an empty Vector Search index, which Vertex AI doesn't allow. The data ingestion pipeline requires the index to be deployed first, creating a chicken-and-egg situation. Also, I had to fix the Makefile to use proper terraform outputs instead of display names

• Issue: Following the documented workflow in README.md fails
• Commands: make setup-dev-env followed by make data-ingestion
• Problem: Chicken-and-egg issue with Vector Search deployment
• Template Version: ADK v1.6.0

Relevant log output

Code of Conduct

• I agree to follow this project's Code of Conduct

[onestardao]

Yup, classic chicken-and-egg trap with RAG index pipelines 🐣
We’ve hit this in a few projects too — Vertex AI being strict about empty index creation really forces a rethink.

One workaround we’ve used:

• Generate a tiny starter corpus (e.g. 2–3 short dummy docs)
• Inject that as “preheat” vector entries to make the index valid
• Once ingestion finishes, overwrite them via reindex or replace call

It’s not pretty but gets the infra past the bootstrapping phase.

Would be cool if future templates handled this edge case by default — maybe an env flag like PRELOAD_INDEX=true to conditionally insert warmup vectors?

btw I’ve compiled 13 recurring AI infra pitfalls like this one (with fixes).
Not dropping them here — but if useful, I can post later.

[luke-swarm-hub]

Yup, classic chicken-and-egg trap with RAG index pipelines 🐣 We’ve hit this in a few projects too — Vertex AI being strict about empty index creation really forces a rethink.

One workaround we’ve used:

• Generate a tiny starter corpus (e.g. 2–3 short dummy docs)
• Inject that as “preheat” vector entries to make the index valid
• Once ingestion finishes, overwrite them via reindex or replace call

It’s not pretty but gets the infra past the bootstrapping phase.

Would be cool if future templates handled this edge case by default — maybe an env flag like PRELOAD_INDEX=true to conditionally insert warmup vectors?

btw I’ve compiled 13 recurring AI infra pitfalls like this one (with fixes). Not dropping them here — but if useful, I can post later.

Please share those 13 recurring AI infra pitfalls, that would be really helpful. Are you able to share directly with me?

[onestardao]

Hey Luke — this is such a classic RAG landmine it should probably come with a warning label 😅
We’ve run into the same chicken-and-egg trap in multiple deployments. Glad someone finally documented it cleanly!

Turns out this issue touches at least three recurring AI infra failure modes we’ve tracked:

1 Hallucination & Chunk Drift: bootstrap ingestion on empty vectors will inject junk.
6 Logic Collapse: pipeline logic assumes index exists before it does.
8 Debugging is a Black Box: infra failures offer zero semantic traceability (why is this failing? no idea).

We actually built a full Problem Map just to catalog stuff like this across hundreds of RAG setups —
all open-source, and fixes included.

Map: https://github.com/onestardao/WFGY/blob/main/ProblemMap/README.md
This specific trap:

• hallucination.md
• logic-collapse.md
• retrieval-traceability.md

The tool that powers all this is open-source (core engine + prompt debugger + logic sanitizer).
If it helps you tame the chaos — feel free to ⭐ the repo! We built it exactly for this kind of mess 😄

[luke-swarm-hub]

Hey Luke — this is such a classic RAG landmine it should probably come with a warning label 😅 We’ve run into the same chicken-and-egg trap in multiple deployments. Glad someone finally documented it cleanly!

Turns out this issue touches at least three recurring AI infra failure modes we’ve tracked:

1 Hallucination & Chunk Drift: bootstrap ingestion on empty vectors will inject junk. 6 Logic Collapse: pipeline logic assumes index exists before it does. 8 Debugging is a Black Box: infra failures offer zero semantic traceability (why is this failing? no idea).

We actually built a full Problem Map just to catalog stuff like this across hundreds of RAG setups — all open-source, and fixes included.

Map: https://github.com/onestardao/WFGY/blob/main/ProblemMap/README.md This specific trap:

• hallucination.md
• logic-collapse.md
• retrieval-traceability.md

The tool that powers all this is open-source (core engine + prompt debugger + logic sanitizer). If it helps you tame the chaos — feel free to ⭐ the repo! We built it exactly for this kind of mess 😄

None of the linked documents actually address deployment ordering or bootstrap ingestion issues. They're all about post-deployment operational problems

[onestardao]

Great catch — you're absolutely right this one slipped past our original scope!

Most of our current docs (like hallucination.md, logic-collapse.md) focus on post-deployment failure states — we hadn’t yet broken out a category for bootstrap/infra ordering failures like this one.

We’ll add a new entry to the ProblemMap specifically for:

"Pre-deployment logic errors (e.g., vector index assumed to exist before it's deployed)"

Thanks again for the correction — your example helps plug a real blind spot. The whole goal of WFGY is to build up a total map of these messy edge cases — and this one definitely deserves its own slot.

If you're okay with it, we might even cite this example (credited) in that new entry.

Map updates soon. 🙏