Loading…
Custom RAG Development · Minneapolis
Minneapolis has the highest concentration of Fortune 500 headquarters per capita in the country: UnitedHealth Group in Minnetonka, Target downtown, 3M in Maplewood, Cargill in Wayzata, Best Buy in Richfield, US Bank in the IDS Center, Medtronic just across the Mississippi, plus Ecolab, Securian, Polaris, and Allianz Life. The document corpora at these companies are the kind that no off-the-shelf search product is going to handle well.
A UnitedHealth Optum analyst at $120,000 fully loaded who spends 25 percent of their time on document search burns $30,000 a year in retrieval time. Multiply by a 50-analyst department and a custom RAG system pays back in months. The same math holds for a 3M R&D group querying lab notebooks, a Target merchandising team searching supplier contracts, or a Medtronic regulatory affairs team cross-referencing 510(k) submissions.
Custom RAG is scoped and quoted as a fixed engagement. Build vs Glean vs Copilot is a real decision and we walk through the three-year TCO during discovery.
Tell us about the corpus and the query patterns.
Glean at $40 per user per month and Microsoft 365 Copilot at roughly $30 per E3 user per month look reasonable on the surface. For an 80-analyst department, Glean is about $38,000 per year ongoing. Over three years that is $115,000 before counting the change-management cost of standing up a new search tool across the team.
A custom RAG built around a specific corpus and a specific set of query patterns is a one-time build, with a few hundred dollars a month in run-rate for embeddings, LLM calls, and vector storage. The trade-off is real: a bespoke system covers exactly what you need rather than a broad SaaS surface area.
We do not pitch custom builds when off-the-shelf is the right call. For a sales team that needs to search Salesforce, Confluence, and Google Drive together, Glean is usually correct. For an Optum medical-policy team or a 3M lab-notebook group, custom usually wins on coverage, latency, and cost.
The architecture we ship for Fortune 500 corpora is conservative on purpose. Vector storage on pgvector inside an existing Aurora cluster (or OpenSearch with vector engine when the team is AWS-native) so retrieval lives in infrastructure the SRE team already operates. Embeddings from OpenAI text-embedding-3-large or Cohere Embed v4 depending on whether multilingual coverage matters.
Retrieval uses a hybrid pipeline: BM25 keyword for exact phrase recall, dense embeddings for semantic recall, metadata pre-filtering on document type and date, and a Cohere Rerank pass on the top 50 to pick the final top 5 chunks. The LLM layer is Claude Sonnet 5 or GPT-5 through your existing enterprise BAA or DPA, never the public consumer APIs.
Evaluation runs before any production deployment. A labeled set of 100 to 200 real questions from your analysts, with recall@5, MRR, and answer-grounding scores tracked continuously. The harness ships with the build so a future model swap can be validated before it touches users.
Six components that show up in nearly every Fortune 500 RAG engagement we ship for the Twin Cities ecosystem.
Veeva Vault QualityDocs, SharePoint, legacy DMS, Confluence, USPTO XML, FDA eSTAR sections. Each source ingested through a dedicated path with the right OCR, table extraction, and metadata mapping.
OpenAI text-embedding-3-large for general English, BioBERT/PubMedBERT for clinical and pharma vocabulary, chemistry-tuned models for 3M-style materials research. Selection made on benchmark scores from your actual corpus.
BM25 plus dense embeddings, metadata filters on document type, date, business unit, or product code before the vector search. Reduces irrelevant recall and keeps latency under 500ms for the typical query.
Re-ranker over the initial top 50 retrieved chunks raises precision 10 to 15 percent on enterprise document sets. The cost is a few cents per query at typical volume.
Every answer includes document name, section heading, page number, and paragraph index. For regulated workflows (FDA submissions, payer contracts) citations include exact predicate or section references for audit.
Labeled question-answer set of 100 to 200 real queries, recall@5, MRR, and answer-grounding scores tracked over time. A future model swap or chunk-size change has to pass before it reaches production.
The Twin Cities Fortune 500 cluster is unusual: it concentrates in healthcare (UnitedHealth, Optum, Medtronic), food and agribusiness (Cargill, General Mills, Ecolab), retail (Target, Best Buy), financial services (US Bank, Ameriprise, Securian, Allianz Life), and diversified manufacturing (3M, Honeywell Aerospace in Plymouth, Polaris). Each of these has a corpus that does not fit the assumptions of generic enterprise search.
Healthcare and medical-device document workflows are where we see the most consistent ROI. Optum's policy and prior-auth libraries, Medtronic's 510(k) submissions, and the clinical-protocol libraries at Mayo Clinic (Rochester is close enough that we treat it as part of the metro) all benefit from custom retrieval. Retail and supplier-contract corpora at Target and Best Buy benefit from the metadata-filtering and structured-extraction patterns we build.
We work remotely with Twin Cities clients on Central time. Discovery calls run between 9am and 5pm CT. Codebase access is through a deploy key or repo collaborator invite. Code ships to repos you own with full IaC.
Industries where we see strongest fit: healthcare insurance ops at Optum-style scale, medical-device regulatory affairs, retail supplier-contract operations, agribusiness commodity research, manufacturing R&D and quality, and mutual-fund and insurance compliance.
Describe the corpus, the source systems, and the query patterns. We'll reply within one business day with a rough scope, a TCO comparison against the SaaS alternatives, and a price range.