RAG at Small Business Scale

The problem

Acme Manufacturing has 50 employees, 200+ SOP documents, and one person who’d been there 15 years and “just knew” how everything worked. When she retired, production slowed. New hires took 3 weeks to ramp up. People kept asking the same questions in Slack.

They needed a way to answer “how do I calibrate the CNC machine?” without finding the right PDF from 2019.

The solution

An internal knowledge agent that indexes all their documents and answers questions in Slack. Built with reaatech/hybrid-rag, running on their existing infrastructure.

What they did

Day 1: Setup and ingestion

1. Cloned the repo and ran pnpm install
2. Connected their Google Drive (where all SOPs lived)
3. Ran the ingestion pipeline — indexed 200+ documents in about 20 minutes
4. Tested queries: “How do I handle a batch that fails QC?” — got a cited answer with a link to the source document

Day 2: Slack integration and testing

1. Created a Slack bot and connected it to the agent
2. Added 3 people to a test channel
3. Asked it real questions for 2 hours
4. Fixed two things:
   • Added a few missing acronyms to the FAQ
   • Tuned the chunk size down for their shorter documents

By end of day 2, it was live in the general channel.

What they got

Before:

• “Hey anyone know the calibration procedure for the Series 400?”
• “Check the 2019 manual… or ask Dave”
• Hours of back-and-forth, interrupted work

After:

• Same question typed in Slack
• Agent replies in 3-5 seconds with the procedure, a link to the source, and notes about which version applies

Real numbers

• 200+ documents indexed
• ~150 queries per week
• Average response time: 3.2 seconds
• Cost: ~$60/month (LLM API + vector store)
• Setup time: 2 days by one person
• New hire ramp time: 3 weeks → 1 week

What they learned

Start with the noisy questions. Don’t try to index everything. Start with the 50 most-asked questions and build from there.

Chunk size matters more than you think. Their SOPs are short (1-3 pages). Large chunks diluted the answers. Smaller chunks with overlap worked better.

Citations build trust. Every answer links back to the source document. When people can verify, they trust the system faster.

The 80/20 rule applies. The agent handles 80% of questions perfectly. The remaining 20% — niche edge cases, contradictory documents, questions that require judgment — still need a human. That’s fine. 80% is a huge win.

When this works, and when it doesn’t

Works well when:

• You have written documentation (even if scattered)
• Questions are repeated enough to be worth indexing
• Answers are factual, not judgment calls

Doesn’t work well when:

• Everything is in people’s heads (you need to write things down first)
• Questions require real-time data (this is document QA, not a monitoring system)
• You need 100% accuracy with zero human review

The build-vs-buy question

If you have someone comfortable with a terminal and a Google Cloud console, the DIY tier is a weekend project. You own the pipeline, your data never leaves your infrastructure, and you pay only for API usage.

If you’d rather focus on your business and have us handle the setup, training, and tuning — book a conversation. We’ll scope it, build it, and hand it off.