GitHub user Xiao-zhen-Liu edited a comment on the discussion: Task ideas for
the dkNet-AI · Apache Texera Agent Hackathon
# AI Workflows in Texera: LLM, RAG & Agentic DAG
**Pitch.** Today Texera can run ML inference but has *no first-class LLM, RAG,
or agentic-DAG operators* — the only AI surface is the workflow-editing chat in
`agent-service`. We add drag-and-drop AI building blocks so users compose RAG
and agent workflows in the same visual canvas they already use for ETL.
**What we'll ship**
- **LLM operators** — `LLMChat`, `LLMEmbedding`. Reuse the existing
`agent-service` LLM gateway (`LLM_ENDPOINT`, OpenAI-compatible) — no new keys,
no new infra.
- **RAG operators** — `DocumentChunker`, `VectorStoreSink`,
`VectorStoreRetriever` on top of a pluggable Python `VectorStore` abstraction.
Ship **Chroma** in embedded mode (zero-deploy MVP) as the first backend.
- **Agentic primitive** — `Agent` operator: a single visual node containing a
ReAct loop with declarative tools (calculator / web-search / SQL / Python). No
DAG-cycle changes needed.
- **Stretch** — a `SubWorkflow` operator that runs a saved workflow as a child
of the parent execution → enables agents-as-reusable-workflows.
**Why it's a good hackathon fit**
- Most of the work follows the established `PythonOperatorDescriptor` +
HuggingFace pattern → fast to land, low review risk.
- Frontend palette auto-updates from operator metadata → zero Angular work for
the demo.
- Each operator is independently demoable on day one.
**Demo storyboard**
1. *Phase-1 demo:* `CSV(reviews)` → `LLMChat("summarize")` → results panel.
2. *Phase-2 demo:* ingest PDFs → chunk → embed → vector-store; ask question →
retrieve → LLM answers with citations.
3. *Phase-3 demo:* one `Agent` node with `WebSearch` + `Calculator` plans a
trip end-to-end; console streams the ReAct trace.
**Impact.** Turns Texera into a viable visual environment for RAG pipelines and
agent prototypes — opening data-science / education use cases (UCI courses,
research demos) that currently need LangChain glue code.
GitHub link:
https://github.com/apache/texera/discussions/5059#discussioncomment-16924278
----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]
