Gemini 3.8 Flash and the Runbook Router Every Office Uses

01 · Where Office Knowledge Actually Stalls

Most multi-office engineering organisations do not suffer from a lack of documentation. They suffer from documentation that lives inside one office's habits. The deployment runbook belongs to the platform group. The incident postmortems sit in a chat channel two time zones away. The escalation playbook exists only as the memory of the one senior engineer who always takes the call. By the time any of that is written down, it has already gone stale.

Treat internal knowledge sharing as an inference problem instead of a filing problem, and the shape of the work changes. Instead of asking people to maintain pages that nobody rereads, you build a system that answers questions from material your teams already produce: design notes, tickets, runbooks, release notes, postmortems, meeting summaries. A language model becomes the interface between the question and the document pile. The model does not have to be clever on its own; it has to be grounded in what your teams wrote.

This is where a pay-as-you-go API relay such as Quanta earns its place on a team running sustained load. One balance covers the major model families, it works with Claude Code, Codex, Gemini CLI and any OpenAI SDK, and every call is metered on real tokens with cache writes and cache reads priced separately. Heavy volume becomes a line item you can explain to finance rather than a surprise at the end of the month.

02 · The Query Inventory Before the Knowledge Base

Before you choose a model or design an index, collect the questions. Ask each office to save one week of the questions it asks in chat, then group them by intent. You are not building a search engine; you are building an answer path for the same half-dozen question shapes that repeat across every office: how do I deploy this, what changed in the last release, who owns this service, what did we decide about that incident, and which environment variable is safe to rotate.

Write the answer shape for each question type before you write a prompt. A runbook question wants ordered steps and a named owner. A decision question wants the date of the decision and the trade-off that lost. A status question wants the source document and its last update. This small discipline does more for answer quality than any amount of prompt decoration, because it tells the model what a complete answer looks like for your organisation.

• › Deployment questions: steps, rollback path, owner, environment names.
• › Ownership questions: team, escalation route, on-call window, source ticket.
• › Decision questions: the date, the options considered, the trade-off that lost.
• › Drift questions: what changed recently and which document is now stale.

Once the inventory exists, retrieval gets easier. You can tell which answers must cite a document, which must summarise several, and which should refuse and point to a human. Refusal paths matter more than most leads expect: an answer that confidently invents a rollback step costs more than a question that goes unanswered.

03 · Route Retrieval to Gemini 3.8 Flash and Synthesis to Claude Sonnet 5.5

Two jobs are hiding inside one question. The first is retrieval and filtering: reading the query, pulling candidate passages, ranking them, rejecting noise. The second is synthesis: turning those passages into an answer a new hire in another office can follow without context. Those jobs do not need the same model, and splitting them is the single most reliable way to keep a heavy knowledge workload affordable.

Use a fast, inexpensive tier for the retrieval stage. Gemini 3.8 Flash handles high-volume classification, keyword expansion and first-pass summarisation well, which is exactly what retrieval traffic looks like: short inputs, many calls, low stakes per call. Then send only the surviving passages to a stronger reasoning tier such as Claude Sonnet 5.5 for the final answer, where step ordering, hedging and source discipline actually matter.

Keep the split explicit in your application code. Retrieval runs on every question. Synthesis runs only when the retrieval stage returns passages above a confidence floor. Below that floor, the system returns a pointer to the owning team instead of a guess. This one rule usually removes the most expensive failure mode in internal knowledge systems: long, fluent, wrong answers written at premium token rates.

04 · Cache Writes and Cache Reads Change the Cost of Repeated Questions

Internal knowledge traffic is repetitive by nature. The same runbook prefix, the same policy text, the same list of service owners appears in thousands of calls across a quarter. If your billing model separates cache writes from cache reads, as Quanta's does, that repetition turns from a cost centre into a discount.

The design implication is simple and easy to get wrong: put stable material at the front of the prompt and volatile material at the end. Organisation policies, team rosters, service catalogues and runbook bodies are stable; the user's question and the retrieved passages are volatile. A prompt built in that order produces a cache entry that survives across offices, across shifts and across weeks, so the second office asking the same question pays for cache reads instead of a fresh write.

Measure it per office. Track cache write tokens against cache read tokens for each question type, and watch the ratio move as your prompt layout settles. A knowledge system that starts expensive and becomes cheap is healthy; one that stays expensive is usually re-writing its context on every call because a timestamp or a rotating identifier crept into the stable prefix.

05 · A Rollout Sequence That Keeps Ticket Volume Flat

Team leads fail at knowledge rollouts for operational reasons, not technical ones. The system works, the answers are reasonable, and then every office dumps its entire backlog into it on day one while the owning team drowns in follow-ups. A steadier sequence protects the people who have to maintain the thing.

• › Pilot with one question type: pick deployment questions only, from one office, with a named reviewer who checks sampled answers each week.
• › Add a second office once the answer format is stable, so you learn whether the wording travels across locations instead of assuming it does.
• › Open the full query inventory only after refusal paths have been exercised and the owning teams know how escalations reach them.
• › Review the metering weekly during rollout, because token behaviour shifts when new question types arrive.

Give each office a local champion who owns its documents rather than a central knowledge team that owns everything. Local ownership keeps sources fresh, and freshness is the constraint that no model can fix. Log which documents get cited, which get flagged as stale, and which questions end in a refusal; those three signals are your maintenance queue.

Budgeting becomes straightforward once the split between retrieval and synthesis is visible, because you can see which question types dominate volume and which ones need a stronger tier. A relay that meters real tokens separately for cache writes and cache reads gives team leads the granularity to make that call week by week, without renegotiating anything when sustained load arrives. If that matches the workload your offices are carrying, open a Quanta account and route your first question type through it this sprint.