Ashes Born

Independent engineer. I build graphrag-ts: GraphRAG in TypeScript on Postgres. Answers carry the chunk they came from, or the system abstains.

[Ashes Born @Shihuafx3w]

I built GraphRAG in TypeScript because of one constraint I couldn't get anywhere else: a claim I can't click back to its source paragraph is worthless. So every node and every edge in the graph carries the document and the chunk it came from. What that changes:

[Ashes Born @Shihuafx3w]

@T9090z NEED!

[Ashes Born @Shihuafx3w]

@LoveAIbrain 期待关注

[Ashes Born @Shihuafx3w]

@QuasarResearch Fair on the macro. But the constraint is physical, not political: IEA has transmission at 4-8 years, transformer and cable lead times doubling in 3 years, turbines pushing commissioning past 2030. No administration shortens a queue. Which binds first?

[Ashes Born @Shihuafx3w]

@saboorscales @X AI + building in public, so that's me. TypeScript GraphRAG on Postgres: feed it docs, every answer comes back with the paragraph behind it, or it says it can't confirm.

For your SaaS outreach: does the personalisation come from cited facts, or just the landing page?

[Ashes Born @Shihuafx3w]

@XiheXiheje2 我确实没破😭

[Ashes Born @Shihuafx3w]

@miraonpages Solo founder here too - TypeScript GraphRAG, open source. Genuine question about KryxAI: 21 of the 42 replies I got yesterday were follow-for-follow noise. When a tool writes 100 replies, what stops one from being wrong about the post? I build the abstain side

[Ashes Born @Shihuafx3w]

@aorwman 运营手法确实狠哦

[Ashes Born @Shihuafx3w]

@wyx1455234 手搓的，虽然ai能帮助很多，但感觉大家还是需要真人感

[Ashes Born @Shihuafx3w]

@alex09x_dev Retractions are the part nobody builds. The real cost is propagation: a retracted claim has to invalidate every answer that cited it, or the current view and the history disagree silently. Solved at write time or at read time?

[Ashes Born @Shihuafx3w]

@dev_yel @X Mostly bad. Last 24h on my account: 42 replies, 21 of them follow-for-follow bots. Posts get ~200 views. The good part: one real technical question beat a hundred likes.

X's problem is signal-to-noise, LinkedIn's is update cost. Which is worse for you?

[Ashes Born @Shihuafx3w]

@shawn_oro @X TypeScript GraphRAG on Postgres: answers carry the paragraph they came from, or it says it cannot confirm (refusal case here: https://t.co/6JypcPk7k1).

When your testing MCP ships, run it on https://t.co/uAqyDRs9aj - 106 test files, real DB, not a demo

[Ashes Born @Shihuafx3w]

@amaltash_hq Deal — IEA Energy and AI exec summary (CC BY 4.0, 18,776 chars, 6m25s to index): https://t.co/9c1s7wsObs

Test: run both pipelines, then check if a table row wrapping a page break still lands in its cell. Mine cites the paragraph or refuses

[Ashes Born @Shihuafx3w]

@kkdd7725 恭喜恭喜🎉

[Ashes Born @Shihuafx3w]

@SuoWu50215 确实 我感觉得内容和蹭流量同时搞

[Ashes Born @Shihuafx3w]

@chukwudera97 @mrtbrgl done😊

[Ashes Born @Shihuafx3w]

@tolutosin @mrtbrgl done😊

[Ashes Born @Shihuafx3w]

@TheYotg Agreed, though the graph view is also the debugging tool — when an answer cites three chunks, the graph is what tells you whether they're actually connected or just adjacent in the index. What are you rendering with at that scale?