Agentic RAG · Magic Pulpo
Answers from your data, not from the model's memory.
Agentic RAG systems that call your database through typed tools, retrieve from a corpus you own, and are measured against a question set before anyone scales them. We are building one for Dosemojo, our own product, on the same standard.
Our point of view
Demos impress. Production has to hold.
- Knows which questions it must refuse
- Scores retrieval and answers separately
- Has a named owner when it drifts
The standard
How a RAG system earns trust.
01
Tools first, retrieval second
Structured questions go to your database through typed tools, not to a similarity search: "grinders under 200 euro" is SQL, not cosine distance. Retrieval runs only on the turns that need it, which is cheaper and more accurate than stuffing every prompt with context.
02
Hybrid retrieval, then a reranker
Keyword search and vector search fused into one ranking, because exact terms matter as much as meaning. One embedding model across documents, queries and jobs, an index sized for the corpus, follow-up questions rewritten before they are embedded, and a reranker once the corpus is large enough to need one.
03
Evaluated before it scales
A golden set of questions with the sources that should answer them, runnable as a script. Retrieval and answer quality are scored separately, so you know which half is failing. No corpus work gets funded until those numbers move.
The engagement
How it works
A structured engagement from first call to measurable results.
- 01
Use cases and guardrails
Which questions the system answers, which it must refuse, and where a person approves before anything leaves the building. Written down before any code, because it decides the architecture.
- 02
Corpus and permissions
What you own or license, how it is chunked, and who may see which part. Access rules are enforced in retrieval, not asked for in the prompt. We do not ingest scraped or pirated sources: that is a copyright case waiting to happen, and the model already knows the public material.
- 03
Retrieval build
One pinned embedding model everywhere. Keyword and vector search fused, an index built for the corpus size, conversation-aware query rewriting, and reranking on top. Tiers and permissions are filters, not score multipliers.
- 04
Eval harness
Thirty to fifty golden questions with expected sources, running as a script in CI. It catches the silent failures: a changed embedding model, a broken ingest job, an index that was never built.
- 05
Production hardening
Authenticated endpoints and per-user rate limits, prompt caching, scheduled ingest and embedding jobs, dashboards for cost and latency, and a deterministic fallback for every agent that can time out.
What you get
Every engagement is scoped and priced upfront. No retainer surprises.
- 01Agent tool definitions with typed inputs and outputs, in your repo
- 02Hybrid retrieval: keyword and vector search fused, with reranking
- 03One pinned embedding model, and the migration path for when it changes
- 04Vector index and schema migrations, versioned with the application
- 05Golden-question eval set and the script that runs it in CI
- 06Guardrail policy: what is refused, what escalates, who approves
- 07Cost, latency and answer-quality dashboards
- 08Scheduled ingestion and embedding jobs, with no manual step
- 09Runbook and a handover session with your engineers
What this costs
Indicative: the final range is set on the call. Fees only; media spend is never marked up. Figures in EUR.
We run this ourselves
We are building one for Dosemojo.
Dosemojo is our own coffee product. Its assistant answers from a curated catalogue and each user's gear, stash and brew history: structured questions go to SQL through tools, the rest to hybrid retrieval over a corpus we own or license, with specialist agents that fall back to deterministic defaults when they run long. The eval set comes before the corpus work, for the same reason we would tell you to build one.
See the ventureThe proof
