Knowledge-base agents are the use case that looks easy and fails hardest. Off-the-shelf RAG returns relevant-looking chunks. The first time finance asks a legal-policy question and gets engineering's answer, trust collapses. GoodMem makes the difference between an agent that ships and an agent that gets shut down by compliance.
The pain, in the client's words
"Our knowledge base has fifteen years of policy documents, support cases, contract templates, and engineering wikis. The agents we built return three different answers to the same question depending on which corner of the corpus they hit. Compliance is asking us to either fix it or shut it down."
Why standard RAG falls short
Off-the-shelf RAG returns chunks that look relevant. It doesn't know which document is current, which is superseded, which one this user is allowed to see, or which one applies to this product line. The agent is confident when it shouldn't be. Customers ship the demo, ship the agent, and lose trust the first time the answer is wrong on a question the company actually cares about.
What GoodMem does differently
- Metadata-aware retrieval. Documents are tagged at ingest with effective date, owner, audience, product line. Queries filter against that metadata so the agent never surfaces a superseded policy or a document the requester can't see.
- Embedder and reranker tuning. Knowledge-base vocabulary is domain-specific. Agent Tuner fine-tunes embedding and reranking on the client's actual corpus and actual question patterns.
- Confidence-gated answers. Confidence scores are surfaced. Below threshold, the agent says "I'm not sure" and routes to a human. Above threshold, it answers with citations. Compliance gets the answer they actually want.
See it live
We built a working knowledge-base agent on top of SailPoint's public documentation, support content, and product references. Multiple corners of the corpus, multiple product lines, ambiguous questions. GoodMem retrieval, GoodMem ranking, GoodMem verification.
This is the kind of agent a Slalom Account Manager can demo to a client cold. Same architecture works for an internal HR knowledge bot, a customer-support FAQ, a legal-policy search agent, or an entitlement-review assistant.
Where this lands inside Slalom
Customer experience and support practices (high-volume FAQ, deflection metrics). Compliance and legal (policy search, contract Q&A, regulatory query agents). Internal IT and knowledge ops. Cyber and identity (the SailPoint pattern: entitlement reviews, policy authoring).
How the engagement is shaped
30 to 60-day pilot scoped to one corpus and one user audience. Measurable target: deflection rate, accuracy on a held-out test set, escalation rate. If the numbers hit, production rollout converts to a 12- or 24-month GoodMem license plus Slalom services on integration, governance, and corpus expansion.