building: @reopenly_ai 🚀 prev: @LandingAI
RJ, Brasil
Joined September 2023
- Tweets280
- Following368
- Followers228
- Likes1.2K
Pinned Tweet
RAG results still bad? Your embedding model may be too generic. Tuning it helps: up to 41% better retrieval in Google Cloud's tests (12% on average).
New on ReOpenly: your server reports each search, you keep what helped, and you fine-tune and host your own model.
How it works: reopenly.com/blog/embedding-…
Why search and RAG fail:
Similar is not relevant. The near miss.
A generic embedding model misses the words of your domain.
Why a search for "the state of the art" could fail?
What if "state of the art" isn't the right term for your domain?
If you want to search your knowledge base well, you need to handle those cases too.
A generic embedding model misses the words of your domain, it will not help you to tune what you want
Here is how to do it in ReOpenly:
1. Report what you are doing
2. Review the results
3. Train a model on your data
4. Compare with the base model
Google Cloud measured up to 41% better search results after tuning text embeddings, 12% on average, on public search benchmarks docs.cloud.google.com/vertex…
How you split documents matters. A March 2026 study tested 36 ways to chunk them in six domains. Paragraph groups put a right result on top about 24% of the time; fixed-size character chunks, about 2 to 3%. Bigger embedding models scored higher, but bad splits still hurt them.
Look at your chunks before you blame the model.
The study, Shaukat, Adnan and Kuhn (March 2026): arxiv.org/abs/2603.06976
Keyword search beat embeddings on financial documents. In a 2026 benchmark of 23,088 questions, keyword search (BM25) got more of the right documents into the top 5 (64.4%) than OpenAI's text-embedding-3-large (58.7%). Both together: 69.5%. With a reranker on top: 81.6%.
Generic embeddings can miss exact terms like company names and fiscal periods. Test yours against plain keyword search on your own questions.
The study, Akarsu, Karaman and Mierbach (April 2026): arxiv.org/abs/2604.01733
For all the people building search
Have you noticed that a generic embedding model misses the words of your domain?
Try this first with free credits, before you train anything:
1. Similarity: your question vs three of your passages
2. Vector: six texts, see which ones group together
3. Report 20 real searches (free), review a few
Train when you have reviewed searches.
The same three in code, one command each, and a comparison with the base model on searches you kept aside: github.com/reopenly/cookbook
The guide: reopenly.com/blog/embedding-…
What is a hard negative in the RAG context, in four posts 📝
1/ A hard negative is a result that looks right and is not: your search ranked it near the top, and it did not answer the question. "How long do refunds take?" answered with the return policy.
Sources: Google Cloud docs.cloud.google.com/vertex… · NVIDIA huggingface.co/blog/nvidia/d… · Databricks databricks.com/blog/improvin…
The guide: reopenly.com/blog/embedding-…
Reporting searches is free. Limits: up to 20 results per reported search; up to 512 tokens per input, 64 inputs per request.
ChatGPT Ads is now available in 26 more countries and 100k+ advertisers have already used it.
But what does this mean for people building on top of these models?
People will take decision based on ads
openai.com/index/chatgpt-ads…
