Updated
Why do my team's AI agents give inconsistent answers?
Because each agent retrieves from a different slice of the repo, and similarity search ranks text by relevance, not by which copy is current, so two agents can confidently cite two stale docs. The fix is a shared source of truth with a freshness signal: one canonical doc per question, marked current. trovex serves that to every agent over MCP, so the team converges instead of drifting.
Same question, different stale doc
When several markdown files touch the same topic, a runbook, an old ADR, a wiki page, each agent retrieves whichever scores highest for its query. Similarity ranks the closest-looking text, not the current one, so one agent reads last year's doc and another reads this week's, and both sound certain. Across a team that reads as the agents disagreeing, when really they are each reasoning over a different copy of the truth.
The fix is a shared source of truth, not a better model
trovex indexes your docs and serves the single canonical doc for a question, marked current, with stale and duplicate copies flagged so they never reach the agent. Every client on the team (Claude Code, Cursor, Windsurf, Cline) reads the same answer over MCP, so the fleet converges. It runs locally with a local embedder and no API key, and an agent can write back what it learns so the next one reads it instead of re-deriving a different version.
Where trovex fits, and where the consulting does
trovex gives the team one consistent source of truth, and you can install it today. If the inconsistency is part of a wider operating problem, how the team runs agents in production, that is what tsukumo consults on. Here is what getting help running agent fleets in production looks like.
FAQ
What is the root cause of inconsistent agent answers?
Each agent retrieves from a different slice of the repo, and similarity search ranks text by relevance, not by which copy is current. So one agent reads an old runbook and another reads the new one, and both answer confidently. Without a freshness signal, the fleet has no shared notion of which doc is the source of truth.
Will a bigger or better model fix the inconsistency?
No. The model is not the problem; the inputs are. If two agents read two different stale docs, a smarter model just reasons more confidently over the wrong one. The fix is upstream: serve every agent the same current doc, not a better guesser.
How does a freshness marker keep agents consistent?
trovex marks each doc canonical, stale, or duplicate and serves only the current one for a query. Every agent on the team gets the same answer for the same question, so they converge instead of each reasoning over a different copy. When a doc changes, the marker moves with it.
Does this work when our docs span multiple repos?
Yes. trovex indexes the markdown you point it at across repos and serves the one canonical answer over MCP, so agents working in different repositories still read from one shared source of truth rather than each repo's local, possibly stale, copy.
Give every agent the same current answer.
Index your markdown and serve the whole team one canonical doc per query, in about a minute.
uv tool install trovex
Open source. No cloud, no API keys. Your docs never leave your machine.