小宋今天忙

只记录不解释

[Arindam Majumder 𝕏 @Arindam_1729]

Local RAG, clearly explained!

(bookmark this)

Most people debug RAG by swapping the model. That's usually the wrong place to look.

Here's the full pipeline I ran on my laptop:

→ Load: 3 plain txt files
→ Split: chunk size 500, overlap 50 (6 chunks total)
→ Embed: nomic-embed-text via Ollama
→ Store: Chroma, fully local
→ Retrieve: top 3 chunks per question
→ Generate: Gemma 4 E4B answers from only those chunks

The key insight: retrieval sets the ceiling, not the model.

If the right sentence never lands in those 3 chunks, no model can give you a grounded answer. A bigger LLM just hallucinates more confidently.

So when an answer is wrong, I check the retrieved chunks first. Nine times out of ten, that's where the problem is.

Local makes this easier. No API cost per test, so you can rerun retrieval as many times as you want.

[小宋今天忙 @Stephanie060109]

@YuFernando 新能源短途代步如果押金和信用免押能打通，接受度会比现在高很多。

[小宋今天忙 @Stephanie060109]

@OPChoudhary_Ind the wind shifted and all the hits died, typical

[小宋今天忙 @Stephanie060109]

@OPChoudhary_Ind lure box has more regrets than successes, but one more

[小宋今天忙 @Stephanie060109]

@n1nikolay 小小一台但辨识度在线，比那些快看不懂脸的新能源强，一眼就记得住

[小宋今天忙 @Stephanie060109]

@emilyyroberta 请问这车适合新手吗？座高友好不？

[小宋今天忙 @Stephanie060109]

@Saudi10557970 166的轴距是真的迷你，感觉停在巷子里都不用倒半天。

[小宋今天忙 @Stephanie060109]

@rhockzy28 试骑报告做成这样比我平时看的参数表有代入感，但要是能加上车型和价格就更完整了。

[小宋今天忙 @Stephanie060109]

阜阳也能安排的话我倒想试试，看看是不是同一个脾气。

[欣然试驾记 @marckkennethdel]

合肥😏今日自行车试骑报告｜这辆平时看着挺规矩，熟悉以后却有点调皮。节奏对上了就特别顺，偶尔还给人一点小惊喜。太老实的怕没意思，太有个性的又让人惦记，真难伺候。芜湖、滁州、阜阳也能安排。更多自行车试骑报告看主页