The brain behind the eyes: How Dosu compares to Glean
Glean revolutionized internal company search but that's where it stops. Dosu does what's next and synthesizes the knowledge your agent finds to provide the context they need to build durable products efficiently.
Self-learning knowledge cache tailored for you and your agents
Dosu improves coding-agent performance by drawing on a knowledge layer that keeps learning. No re-reading the entire codebase on every task — your agents get exactly the context they need, when they need it.
# eng-help
Cloud Filestore (NFS) — a 2.5 TB BASIC_SSD instance mounted at /mnt/filestore. It stores LanceDB vector databases that power code search and retrieval. The internal-api mounts it read-write for indexing, and the search-api mounts it read-only for queries [1].
The search-api Cloud Run service sits in front of it, handling vector and full-text search requests from agents over a direct VPC network interface for low-latency access.
Updated! I've added the Cloud Filestore / LanceDB architecture details to the infrastructure docs. Let me know if anything needs adjusting.
Code Retrieval Infrastructure
Notion
Engineering / Infrastructure / Code Retrieval Infrastructure
Code Retrieval Infrastructure
Last edited 4 min ago
Our code search and retrieval pipeline is backed by Cloud Filestore (NFS) — a 2.5 TB BASIC_SSD instance mounted at /mnt/filestore. It stores LanceDB vector databases that power all vector and full-text search across the codebase.
Source: This page was reconciled from a thread in #eng-help started by Miguel on Tuesday at 4:43 PM.
Filestore configuration
| Property | Value |
|---|---|
| Instance type | BASIC_SSD |
| Capacity | 2.5 TB |
| Mount path | /mnt/filestore |
| Protocol | NFS |
| Network access | Direct VPC (low latency) |
Knowledge your agents learn from and write back to
Search is a one-way street: it reads what already exists and hands it to whoever asked. Whatever an agent works out along the way stays in that one session and disappears with it.
Dosu closes the loop. Answers are written back to the source they came from — across code repos, docs, and team communication tools — so the knowledge layer compounds instead of fragmenting into conflicting copies.
AI-native knowledge for AI-native engineers and agents
Enterprise search was built to help people find documents. Dosu generates and maintains the structured context that coding agents actually consume — reconciled on every pull request, served over MCP and the CLI.
Michael Torres
Staff Engineer
Sarah Chen
Senior Engineer
Jun Park
QA Engineer
Horizontal search and vertical knowledge do different jobs
Glean is horizontal: one search box across everything your company has already written, for every team in the building. Dosu is vertical and engineering-specific: it writes and maintains the architecture decisions, incident context, and system reasoning that were never documented in the first place — the context no index can surface, because it does not exist yet. Teams run both, and neither replaces the other.
You built something
worth documenting
Let Dosu handle the part you've been putting off.