Retrieval Workbench

RAG Signal Forge

Tune retrieval quality like a systems engineer. This lab simulates the search loop used in local-first markdown RAG stacks and lets you stress-test your query strategy.

Embeddings Chunking TopK / Score Tuning Latency Focus

Core Loop

Index

argo-rag index --vault ~/Vaults/knowledge-base

Chunk markdown, generate embeddings, and build vector index.

Search

argo-rag search "docker bridge networking problems"

Run semantic lookup and inspect top chunk/source matches.

Serve

argo-rag web

Open local UI to compare query phrasing and source quality.

Live Container Lab

This route now launches real ephemeral containers. Use them as control/target nodes to rehearse your index-and-query workflow before running against your full vault.

Connecting to lab engine...

Checking if real Linux environments are available

Challenges

Validate your container environment before installing any RAG tooling.

○ Check Python version

○ Verify pip is available

○ Check available disk space

○ Verify network connectivity

○ Check package manager availability

Install and verify the core Python libraries needed for local RAG.

○ Install sentence-transformers

○ Install ChromaDB

○ Install numpy

○ Verify all imports work

○ Check GPU availability for embeddings

Create and organize a markdown corpus for retrieval testing.

○ Create a markdown corpus directory

○ Write sample markdown files

○ Count files in the corpus

○ Inspect frontmatter structure

○ Plan a chunking strategy

Generate embeddings from your corpus and build a searchable vector index.

○ Generate embeddings for a document

○ Build a ChromaDB collection

○ Inspect the index files on disk

○ Measure indexing time

○ Verify embedding dimensions

Run semantic searches and tune retrieval quality.

○ Run a semantic search query

○ Compare top-k result sets

○ Tune score thresholds

○ Measure query latency

○ Test query variations

Connect retrieval to prompt construction and evaluate end-to-end quality.

○ Connect retrieval to a prompt template

○ Test context injection quality

○ Compare answers with and without RAG context

○ Evaluate answer quality manually

○ Optimize chunk size for your corpus

Benchmark and optimize embedding and indexing throughput.

○ Benchmark embedding throughput

○ Compare GPU vs CPU embedding speed

○ Test batch indexing strategies

○ Profile memory usage during indexing

○ Design an embedding cache strategy

Build resilience with recovery procedures, monitoring, and quality checks.

○ Test index corruption recovery

○ Implement incremental index updates

○ Set up monitoring metrics

○ Manage multiple collections

○ Build automated quality checks

0% Complete

Quick Reference

Python / pip

pip install sentence-transformers chromadb python3 -c "import chromadb; print(chromadb.__version__)"

Embedding

from sentence_transformers import SentenceTransformer model = SentenceTransformer('all-MiniLM-L6-v2')

ChromaDB

client = chromadb.PersistentClient(path="./chroma_db") collection.query(query_texts=["search term"], n_results=5)

File Operations

find /corpus -name "*.md" | wc -l du -sh ./chroma_db

Container 1: Control Node

LIVE Connecting...

Container 2: Target Node

LIVE Connecting...

Lab Engine Status

Checking

Active sessions --

Capacity --

Template --

Last check --

Waiting for lab engine response...

Live Retrieval Simulator

Query TopK 3 Min score 0.35

0 results Latency: 0ms

Index Cost Estimator

Markdown files 3000 files GPU acceleration enabled

Estimated index time

~35 min

Estimated query latency

<100 ms

Tuning Missions

0/5 complete

Run 3 semantically different queries for the same issue and compare top hits. Tune topK + minScore until irrelevant chunks stop appearing. Measure index speed with and without GPU and log the delta. Add frontmatter tags to weak docs and rerun retrieval quality checks. Document one prompt template that consistently gives useful answers.