Building a RAG knowledge base for your marketing team
Stop searching Notion. Start asking it.
Most marketing teams have a five-year archive of campaign learnings nobody can find. A retrieval-augmented chatbot trained on that archive changes how the team plans. The first time someone asks 'what did we learn from the spring 2024 launch?' and gets a real answer in 4 seconds, you've justified the build.
The stack we use
Supabase pgvector for storage, OpenAI embeddings, a thin internal app served from Vercel. Total monthly cost under £200 for a 20-person team. Resist the urge to over-engineer with vector databases that need a dedicated cluster — pgvector is plenty for under 500k chunks.
The use cases that drive adoption
- Pre-campaign briefs — pull every related learning into the brief automatically.
- Creative concept screening — has this idea been tried? What happened?
- Post-mortem comparisons — how did this launch perform vs the last three?
- Onboarding — new hires self-serve historical context in week one.
Adoption strategy
Make the first query each new user runs return something they couldn't find before. Without that magic moment, the tool becomes another bookmark. Seed the chatbot with the team's twenty most-frequently asked questions and answers — pre-warm the experience.
Governance
Tag every source document with confidence level and date. The model will happily quote a 2021 channel forecast as if it's current. Filter by recency in the retrieval step, not the prompt.