Track C: RAG and vector memory

Track C walks through document ingestion, vector memory configuration, and query runs that emit MemoryRetrieved trace events. You use Qdrant plus embedding configuration (stub or real provider).

Goal

Ingest documents, query via a vector-memory agent, and verify retrieval in execution traces. First contact with Qdrant, embedding env vars, and hybrid retrieval settings.

Prerequisites

Stub-only path works without Qdrant but will not emit MemoryRetrieved until documents exist in the store.

Step 1: Start Qdrant (recommended)

If your dev compose includes Qdrant, start it with the server stack or standalone container on port 6333. Export:

export ARCFLOW_QDRANT_URL=http://localhost:6333

Step 2: Configure memory on an agent

Use the sample from document_qa.py:

memory = MemoryConfig(
 MemoryType.VECTOR,
 MemoryScope.AGENT,
 namespace="track_c_kb",
 embedding="stub/8", # swap for production embedding id
 retrieval=MemoryRetrievalConfig(
 mode="hybrid",
 dense_weight=0.7,
 sparse_weight=0.3,
 top_k=3,
 ),
 chunking=MemoryChunkingConfig(chunk_size=256, overlap=32),
)
agent = Agent(name="researcher", role="researcher", instructions="Answer using retrieved context.", memory=memory)
workflow = Workflow(name="track-c-rag", agents=[agent])

Change namespace to a dedicated track id to avoid colliding with other tutorials.

Step 3: Ingest sample documents

Ingest text into namespace track_c_kb using admin ingest API, dashboard Knowledge tab, or SDK ingest helpers described in knowledge ingestion.

Minimal text for verification:

ArcFlow hybrid retrieval combines dense vectors with sparse signals.
Chunk size and overlap affect recall on long documents.

Confirm ingest success before querying.

Step 4: Run query

python examples/rag/document_qa.py

Or inline:

result = workflow.run("What does ArcFlow use for hybrid retrieval?")
print(result.output)

Step 5: Verify trace events

kinds = {e.get("event_kind") for e in result.trace_events}
assert "MemoryRetrieved" in kinds, f"missing MemoryRetrieved; got {kinds}"
assert result.status == "completed"
print("track C trace checks passed")

Inspect MemoryRetrieved payloads for chunk counts and scores without chunk text (trace data policy).

Verification checklist

Expected output

Non-empty answer string referencing hybrid retrieval concepts when the store is populated. Verification prints track C trace checks passed.

Troubleshooting

What you learned

Track C connects agent memory configuration to operational ingest and trace-verified retrieval. Platform RAG features in static product and server admin APIs build on the same vector primitives.

Next tracks