Ask in plain English. Get the source back
A team asks a question and gets an accurate answer from its own documents, databases and knowledge. Every answer traceable to where it came from, behind the access controls you already run.
Confidence comes from verifying the evidence, not from the model grading itself.
4
retrieval systems in production behind this practice
6
mechanisms we tune, each by name
0
answers shipped without their source
Four retrieval systems, four different hard parts
Retrieval is not one problem. A portfolio question, a lesson plan, an ERP query and a support conversation each break in a different place.
A temporal knowledge graph with evidence-based answers
For a family-office platform: a graph with valid-from and valid-to on its edges, so a partner can ask what the portfolio looked like on any past date. Every extracted value carries its exact source quote and page, and confidence comes from verifying that the quote exists — not from the model grading itself.
Hybrid retrieval engineered to the decimal
For a K-12 publisher: dense embeddings fused with keyword ranking by reciprocal-rank fusion, a candidate pool of fifty narrowed to ten, progressive filter relaxation across trait, mode and grade, and a five-category source hierarchy so the client's own books always outrank reference material.
Natural language to SQL and API, delivered
For Dyco Parts: plain-English questions answered from an Odoo ERP, e-commerce channels, meeting transcripts and scanned invoices at once — multi-modal GraphRAG, delivered in Slack.
Retrieval at support scale
In VORTA, our own support product: hybrid retrieval over knowledge bases, multi-turn and multilingual, where the cost of a wrong answer is a customer rather than a report.
This audience rewards specificity
So here are the parts we actually tune, rather than a claim that we tune them.
Hybrid search
Dense retrieval catches meaning; keyword retrieval catches the part number. Choosing one is choosing which half of your questions to answer badly.
Reciprocal-rank fusion
The two result lists are merged by rank rather than by score, because their scores are not on the same scale and pretending otherwise quietly favours one.
Progressive relaxation
Filters loosen in a defined order when nothing matches, so a narrow question degrades to a useful answer instead of an empty one.
Temporal edges
Relationships carry validity dates, so the graph can answer what was true then, not only what is true now.
Reranking and semantic caching
A reranker promotes the passage that actually answers the question; caching keeps the repeat questions cheap.
An evaluation harness
Hand-written retrieval test questions, run on every change. Without them, retrieval quality is an opinion.
Retrieval quality is a discipline, not a vibe
On one platform: full classification accuracy on the test set and roughly 97% quote accuracy, with a multimodal change taking scanned-document extraction from about 12% to about 94%.
Those are project test results, not a promise about your corpus. What transfers is the harness that produced them: on another build, 2,092 automated checks run in under ten seconds on every change, with 106 live acceptance checks behind them. If retrieval is not measured, nobody can tell you whether a change made it better.
The systems built on this practice
Bring twenty questions your team cannot answer today
This is our AI engineering practice
It is real work and it is where our four products came from. But what Cognilium leads with is narrower: optimization apps that run in tandem with Microsoft Dynamics 365, computing the decisions the ERP records but does not derive — the optimal price, the optimal pick path, the optimal stock level. See the optimization apps · How we build inside the ERP.