--- title: Do You Need a Custom Embedding Model? | daily.dev description: A decision framework for when to fine-tune your own embedding model instead of using off-the-shelf APIs like OpenAI's text-embedding-3-large or Cohere's... canonical: https://daily.dev/posts/do-you-need-a-custom-embedding-model--0goip7lmx twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Do You Need a Custom Embedding Model? | daily.dev og:description: A decision framework for when to fine-tune your own embedding model instead of using off-the-shelf APIs like OpenAI's text-embedding-3-large or Cohere's... og:url: https://daily.dev/posts/do-you-need-a-custom-embedding-model--0goip7lmx og:image: https://api.daily.dev/og/posts/0gOIp7lMx.png og:image:alt: Do You Need a Custom Embedding Model? og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Do You Need a Custom Embedding Model? **[DigitalOcean Community](https://daily.dev/sources/do_community)** · 11 min read · 1 upvotes · 1 comments ## Summary A decision framework for when to fine-tune your own embedding model instead of using off-the-shelf APIs like OpenAI's text-embedding-3-large or Cohere's embed-v4. The threshold: if recall@10 on domain queries falls below ~80% and you can collect a few thousand real query-to-document pairs, fine-tuning a small open model like bge-base-en-v1.5 typically beats API models by 10+ points at minimal cost (a training run costs roughly $3.40 on an H100 GPU Droplet, under $1 on cheaper hardware). A worked support-ticket RAG example walks through data prep, a 30-line sentence-transformers training script using MultipleNegativesRankingLoss, and evaluation. It also compares infrastructure options (DigitalOcean GPU Droplets, Modal, Baseten, Together AI) and lists three scenarios where fine-tuning isn't worth it: low query volume, fast-moving vocabulary, or lack of labeled pairs. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### When should I fine-tune my own embedding model instead of using OpenAI or Cohere's API for RAG? Fine-tune when recall@10 on your own domain queries falls below about 80% with a general-purpose model and you can collect a few thousand real query-to-document pairs from logs, tickets, or click data. Published fine-tunes of small open models like bge-base-en-v1.5 have beaten OpenAI or Cohere by 10 or more recall points, at a training cost of a few dollars on a single GPU. _Weighing embedding APIs against a custom fine-tune is exactly the kind of RAG tradeoff developers track on daily.dev._ ### How much does it cost to fine-tune the bge-base-en-v1.5 embedding model? Fine-tuning bge-base-en-v1.5, a 109M-parameter model, on 50,000 query-document pairs takes about 40 minutes on a single H100 GPU Droplet, costing roughly $3.40. The same run costs under a dollar on a cheaper card like an RTX 4000 Ada. The real cost of the project is collecting good training pairs, not the GPU time. _Anyone budgeting a RAG fine-tune project can find cost breakdowns like this through daily.dev._ ### How much less pgvector storage does a 768-dimension embedding model use compared to a 3072-dimension model? A 768-dimension model like bge-base-en-v1.5 needs roughly a quarter of the storage of a 3,072-dimension model such as text-embedding-3-large. For 10 million chunks, that is about 31 GB versus 123 GB of vectors plus index in pgvector, a difference that recurs every month and can push a team into a larger database plan. _Engineers sizing pgvector storage for RAG pipelines rely on daily.dev to compare embedding model tradeoffs._ ## Community discussion Top comments from developers on daily.dev. **@kartiknvj** · 0 upvotes > The recall@10 under 80% threshold is a useful trigger to name, most teams jump to fine-tuning long before they've measured that. I've fixed more recall problems with better chunking and a reranker than with a custom model. When you did fine-tune, how much lift did the embeddings actually buy over a reranker? ## Similar posts on daily.dev - [Build a Domain-Specific Embedding Model in Under a Day](https://daily.dev/posts/build-a-domain-specific-embedding-model-in-under-a-day-otroujtin) · Hugging Face · 0 upvotes · 0 comments - [Medium](https://daily.dev/posts/medium-xuslxsabv) · Medium · 0 upvotes · 0 comments - [Rerankers Aren’t Magic Either: When the Cross-Encoder Layer Is Worth the Cost](https://daily.dev/posts/rerankers-aren-t-magic-either-when-the-cross-encoder-layer-is-worth-the-cost-tzsqjvbgp) · Towards Data Science · 0 upvotes · 0 comments - [What RAG Actually Costs to Run in Production: A Full Cost Breakdown](https://daily.dev/posts/what-rag-actually-costs-to-run-in-production-a-full-cost-breakdown-26v3uoonx) · DigitalOcean Community · 0 upvotes · 0 comments --- Tags: [#deep-learning](https://daily.dev/tags/deep-learning), [#rag](https://daily.dev/tags/rag), [#vector-search](https://daily.dev/tags/vector-search), [#embeddings](https://daily.dev/tags/embeddings), [#pgvector](https://daily.dev/tags/pgvector) [View this post on daily.dev](https://daily.dev/posts/do-you-need-a-custom-embedding-model--0goip7lmx) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Do You Need a Custom Embedding Model?","url":"https://daily.dev/posts/do-you-need-a-custom-embedding-model--0goip7lmx","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/do-you-need-a-custom-embedding-model--0goip7lmx"},"datePublished":"2026-08-24T16:02:30.476Z","dateModified":"2026-09-14T07:06:55.130Z","description":"A decision framework for when to fine-tune your own embedding model instead of using off-the-shelf APIs like OpenAI's text-embedding-3-large or Cohere's...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8f2172bba436e9bce87684915dfd91c6?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8f2172bba436e9bce87684915dfd91c6?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"DigitalOcean Community","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"DigitalOcean Community","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/c1b9d07730e34ea388c39a498a753d6c","url":"https://daily.dev/sources/do_community"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/do-you-need-a-custom-embedding-model--0goip7lmx","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"deep-learning,rag,vector-search,embeddings,pgvector","timeRequired":"PT11M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"DigitalOcean Community","item":"https://daily.dev/sources/do_community"},{"@type":"ListItem","position":3,"name":"Do You Need a Custom Embedding Model?"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/do-you-need-a-custom-embedding-model--0goip7lmx","comment":[{"@type":"Comment","text":"The recall@10 under 80% threshold is a useful trigger to name, most teams jump to fine-tuning long before they’ve measured that. I’ve fixed more recall problems with better chunking and a reranker than with a custom model. When you did fine-tune, how much lift did the embeddings actually buy over a reranker?","datePublished":"2026-08-24T18:08:53.281Z","url":"https://daily.dev/posts/0gOIp7lMx#c-K041ZOOQC","author":{"@type":"Person","name":"kartik-nvjk","url":"https://daily.dev/kartiknvj","image":"https://media.daily.dev/image/upload/s--3gGgsVCw--/f_auto/v1781456774/avatars/avatar_TvTVeiMdkRCqWUDullFmy?_a=BAMAMiWQ0"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/do-you-need-a-custom-embedding-model--0goip7lmx#faq","mainEntity":[{"@type":"Question","name":"When should I fine-tune my own embedding model instead of using OpenAI or Cohere's API for RAG?","acceptedAnswer":{"@type":"Answer","text":"Fine-tune when recall@10 on your own domain queries falls below about 80% with a general-purpose model and you can collect a few thousand real query-to-document pairs from logs, tickets, or click data. Published fine-tunes of small open models like bge-base-en-v1.5 have beaten OpenAI or Cohere by 10 or more recall points, at a training cost of a few dollars on a single GPU. Weighing embedding APIs against a custom fine-tune is exactly the kind of RAG tradeoff developers track on daily.dev."}},{"@type":"Question","name":"How much does it cost to fine-tune the bge-base-en-v1.5 embedding model?","acceptedAnswer":{"@type":"Answer","text":"Fine-tuning bge-base-en-v1.5, a 109M-parameter model, on 50,000 query-document pairs takes about 40 minutes on a single H100 GPU Droplet, costing roughly $3.40. The same run costs under a dollar on a cheaper card like an RTX 4000 Ada. The real cost of the project is collecting good training pairs, not the GPU time. Anyone budgeting a RAG fine-tune project can find cost breakdowns like this through daily.dev."}},{"@type":"Question","name":"How much less pgvector storage does a 768-dimension embedding model use compared to a 3072-dimension model?","acceptedAnswer":{"@type":"Answer","text":"A 768-dimension model like bge-base-en-v1.5 needs roughly a quarter of the storage of a 3,072-dimension model such as text-embedding-3-large. For 10 million chunks, that is about 31 GB versus 123 GB of vectors plus index in pgvector, a difference that recurs every month and can push a team into a larger database plan. Engineers sizing pgvector storage for RAG pipelines rely on daily.dev to compare embedding model tradeoffs."}}]} ```