EKOS
Compile enterprise knowledge once, ask LLMs cheaply
13 followers
Compile enterprise knowledge once, ask LLMs cheaply
13 followers
EKOS is an open-source Rust knowledge compiler. It turns schemas and code into an append-only, evidence-backed ledger of tables, relationships and join paths, plus a data profile with nulls, empty values and examples. For text-to-SQL it builds one compact, verified prompt instead of raw DDL, runs the query read-only and retries once on errors or empty results. No agent loop, no extra tools: a small model answers more accurately at a fraction of the token cost. Benchmarked on BIRD Mini-Dev.
Free
Launch Team / Built With
Render Sandboxes: a computer for every agentRun untrusted code safely. Idle sandboxes cost nothing
Promoted
Maker update — distributed storage + a real query engine landed this week 🚀
Two big pieces shipped since launch, both proven against real workloads, not just unit tests:
1. Distributed storage & query is feature-complete (v1). EKOS can now split a knowledge base across partitions and serve reads/writes from independent worker nodes backed by S3/Azure/local object storage — a coordinator hands out write leases, query workers cache and serve partitions, and cross-shard search does a real BM25 top-k merge instead of a naive concat.
We didn't just call it done and move on — we ran two full autonomous end-to-end tests against real 90+ partition workloads (MinIO, live LLM provider, the works). They found 8 real defects unit tests never caught: a runtime panic on the object-store path, a query worker that took down the whole cluster when killed, a heartbeat interval that silently expired write leases mid-pipeline. All 8 are fixed with regression tests, and the reports are public in the repo. If a demo only shows the happy path, it's marketing, not evidence — so we publish the failures too.
2. A real compiled-knowledge query engine, not just "dump some JSON at the LLM." Natural-language questions now compile into a typed query plan, execute against structured facts + graph + search, and hand the model a flat set of atomic, source-cited claims — file, line, extractor, confidence — instead of whole objects for it to re-derive structure from. This is the same "compile once, trust the artifact" philosophy applied to retrieval itself.
Both are on main today. As always — open source, evidence-backed, gaps documented rather than hidden. Happy to answer questions about either piece below 👇
Hey Product Hunt 👋 Maker here with an update on EKOS.
Quick recap: EKOS is an open-source (MIT) knowledge compiler written in Rust. It turns your scattered engineering estate (repos, docs, Confluence, legacy ETL) into a queryable knowledge graph that Claude Code and other agents can use through MCP.
What shipped recently:
🔁 Legacy transformation recovery: `ekos recover` compiles Pentaho/Kettle jobs (.ktr/.kjb), SQL views and stored procedures into one shared Transformation IR. You can now diff old ETL logic against a newly drafted pipeline and see exactly which business rules changed.
🧩 Identity resolution: `ekos identity scan` finds the same table, column or entity under different names across systems, with a review tool so a human confirms the matches.
🛠 New MCP tools for Claude Code: transformation_explain / transformation_diff, impact analysis, dependents, neighborhood search and more. Read-only by default, so it's safe to point an agent at it.
Take a look at:
site https://ekos.dev
Demo Videos:
https://youtu.be/EGpYzn_-fv4?si=fp1lHYNGxfDLSUUi
https://youtu.be/lRjAZkViE2w?si=UM2t-Hd-faAeaOr-