Connectors are the hard part of retrieval: naming the owner

“Connectors are the hard part of retrieval” stops being abstract the moment you ask who gets paged when it fails.

Demo RAG is a weekend. Operable RAG is product work: freshness, permissions, connectors, evals, and an owner who gets paged when answers go stale.

The rest develops that one idea with controls you can install and failure modes you can recognize.

When pressure rises, teams drop the unowned practice first. Put Connectors are the hard part of retrieval in the path of work — templates, checklists, review norms — or admit it was optional applause.

Write one page while the decision is still warm — context, options, choice, owner, revisit date. Verbal alignment on Connectors are the hard part of retrieval evaporates under ordinary calendar pressure, and Slack archaeology is a poor substitute for a decision record.

A scene from delivery

Support escalations mention “the bot was wrong last month too.” Without retrieval traces, engineering debates model temperature instead of the stale source.

That scene is the thesis under calendar pressure. The response is not another status meeting — it is a control.

Mechanisms

Freshness SLOs belong next to product SLOs. A corpus that is “usually fine” is how trust erodes without an incident ticket.

Instrument the pipeline: query → retrieved IDs → scores → answer → user feedback. Make polite lies auditable. If you cannot replay what was retrieved, you cannot improve it.

Ground answers in systems teams actually run — tickets, dashboards, runbooks — not paraphrases of marketing docs. Citation quality is a product requirement.

Partial retrieval should be visible to the user and to ops. Silent degradation is how polite lies scale.

Build evals early — gold questions, regression sets, and failure buckets (wrong, stale, unauthorized, empty). Without them you argue about vibes while users lose trust.

Own freshness: every source needs a refresh policy, a staleness signal, and a human accountable when the corpus lies. Silent drift is how users stop asking.

Resist the urge to expand scope into neighboring slogans. If a control does not make Connectors are the hard part of retrieval more operable for the next person, leave it for another note. Dilution is how coherent essays become stitched scrapbooks.

Another texture

A bot answers from last quarter’s runbooks with perfect tone and wrong on-call. Freshness had no owner. Users learned to distrust the system in a week.

Who owns the knowledge graph behind the bot? “The AI squad” means nobody owns source quality. Assign freshness owners per major corpus with an SLO.

Tradeoffs

Strict authz can reduce recall. That is acceptable. Unauthorized recall is not a quality feature.

Evals feel slow before launch and priceless after. Fifty gold questions beat a thousand vague complaints.

Thin-context refusals lower answer rate and raise trust. Optimize for trust if the product depends on it.

Keep the unit of progress small enough to finish under a full calendar: one owner clarification, one verification signal, one reversible control. Grand programs without weekly evidence become status machines.

For Connectors are the hard part of retrieval, the inheritance test is blunt: after a week, can someone outside the original room explain what changed, who owns it, and how you will know if it breaks? If the answer depends on hallway memory, you still have a story — not a practice.

Failure modes

Each of these is a missing control, not a personality problem:

• Evals skipped because “users will tell us.” Users leave instead.
• Connectors treated as temporary glue with no product owner.
• Shipping the chat UI before connectors and authz — then spending a year apologizing for hallucinations that were access bugs.
• Treating retrieval as a model problem when it is a data-operations problem.

What to do this month

• Review connector failures in the same forum as product incidents.
• Add a thin-context path: refuse or escalate instead of inventing.
• Before UI polish: ship connectors, authz checks, and a 50-question eval set.

When retrieval lies politely, users blame “the AI” and stop reporting. Instrument feedback and retrieval traces so lies become tickets with owners.

Chunking strategy as an architecture choice belongs in the same review as service boundaries. It changes failure modes and cost profiles.

Permissions bugs in RAG are data breaches with a conversational UI. Treat them with the same severity as any other authz defect.

Connectors need product management: identity, rate limits, partial failure, schema drift, and who owns the break. Glue code without an owner becomes a career-limiting outage.

Citation quality is part of the answer. An uncited fluent paragraph trains users to treat guesses as documentation.

Gold-question suites are living artifacts. Retire questions that no longer match the product; add questions from real support failures.

Design chunking as architecture: preserve meaning units, cite sources, and refuse to shred tables into noise. “Improving” chunk size without an eval is a regression with good intentions.

Fail gracefully when context is thin: say what you do not know, ask for a tighter question, or escalate to a human path. Inventing confidence is how demos become liabilities.

Permissions in RAG are not optional. Enforce authz at retrieval time; never “filter in the prompt.” Prompt-side filtering is a leak with extra steps.

Close

Build the eval loop first. Everything else is speculation with a chat box.

On Connectors are the hard part of retrieval, keep the loop short: write the decision, name the owner, verify the outcome.