Updated
trovex for large monorepos
In a large monorepo the agent's bottleneck is finding which of many overlapping docs is current, not reading code. trovex indexes the markdown and returns the single canonical doc per query with a path:line pointer and a freshness marker, instead of letting the agent sift hundreds of files. About 60% fewer tokens per lookup, and the bigger the repo, the more it matters.
The problem at monorepo scale
A big repo has many docs that say overlapping things: an old runbook, a current one, a design note, a wiki export. Similarity search ranks the closest-looking text, not the current one, so the agent opens several and reasons over the pile, every session. The more docs there are, the worse the rereading and the higher the chance it picks a stale one.
One canonical doc, not the whole tree
trovex returns the one doc that answers a query, down to the section, and marks stale and duplicate copies so they are skipped. The agent reads one current answer with a path:line pointer instead of scanning the tree. Lookups stay fast because it returns one doc per query, not a scan of every package. That is what stopping agents picking stale docs takes at scale.
How it rolls out
Point trovex at the markdown across the monorepo and it builds a local index: vectors in SQLite, an on-device embedder, no cloud or keys. Each developer's agent queries it over MCP. Index once, serve every query.
When the docs problem is really an operating problem
A large repo with drifting docs is often a symptom of how the team runs agents, not just a search problem. trovex fixes the retrieval; if making agents reliable across a large codebase is the hard part, tsukumo consults on the operating model, and trovex is the proof.
FAQ
Does indexing a huge repo make lookups slow?
No meaningfully. trovex returns one canonical doc per query rather than scanning every package, and the index runs locally in SQLite, so lookups stay fast as the repo grows.
What does it index in a monorepo?
The markdown you point it at across the tree: docs, runbooks, ADRs, design notes. It returns the one current doc for a query with a path:line pointer and a freshness marker.
How much does it save on a big repo?
On our own repo it measured a median of 69% fewer tokens per lookup at equal task-success across 26 queries, range 41 to 81%; we headline a conservative about 60%. More overlapping docs means more rereading removed. See trovex.dev/measure.
Serve your monorepo's agents the one current doc.
Index the whole tree once; every query returns one canonical answer. About a minute to set up.
uv tool install trovex
Open source. No cloud, no API keys. Your docs never leave your machine.