More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Sentence Transformers: Embeddings That Actually Retrieve

    1

    Pick the right embedding model, encode efficiently, and measure retrieval on real queries — semantic search and RAG that retrieves what users actually need.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Sentence Transformers: Embeddings That Actually Retrieve

    Sentence Transformers: Embeddings That Actually Retrieve

    Example session with this skill installed

    Our support docs search misses the right articles half the time. Pick an embedding model, encode the docs, and show whether retrieval actually improves — measured, not guessed.

    • Read your context and instructions
    • Compiled the sentence transformers

    Collected 30 real support queries from the last month and judged which help article each one should return. Keyword search scored recall at 5 of 0.47 — the baseline the new system had to beat. Prototyped the small generalist first: a 384-dimension model encoded the docs in minutes and reached recall at 5 of 0.71. A 768-dimension model reached 0.76, a real but modest gain that cost roughly twice the storage and latency. Reading the failures mattered more than the average: five misses came from answers split across chunks, and three from product codes the dense model could not match exactly. Chunking was changed to section boundaries with 15 percent overlap, and a keyword layer was added for codes and names. Final numbers: recall at 5 of 0.83 with the 768-dimension model, hybrid search, and section chunking — and the failure list shrank to two genuinely ambiguous queries. The model version is pinned in the config, the embeddings are cached beside their source text with the model name attached, and the re-encoding cost of any future model switch is written down before anyone promises it. The small model stays as the documented fallback: if the larger model ever stops earning its keep on the monthly re-measure, the switch back is one config change, not a rebuild. The evaluation set of 30 queries is saved, so every future model choice is decided by the same evidence instead of a fresh argument.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Most embedding failures are model-selection failures wearing a code costume. This skill covers choosing the right sentence embedding model for the job — small generalists for prototyping, larger models when retrieval is the product, multilingual and domain-specific models only when the data earns them — then encoding wisely: batch, normalize, cache, and chunk by meaning before the model ever sees the text. It builds the retrieval loop that proves the choice: real user queries, judged correct answers, recall measured per candidate, and the failures read instead of averaged away. It also covers when fine-tuning earns its keep, how to keep metadata and provenance with every chunk, and the production rules — pinned model versions, re-encoding budgets, quantization trade-offs, and hybrid search — that keep a working retriever working.

    It teaches the method; it does not encode your documents for you — you run the code in your own environment.
    Recall numbers depend on your queries and your docs; the skill shows how to measure them, it cannot promise a specific score.
    Model recommendations are guidance at the time of writing; verify current model names and sizes against the model hub before committing.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 3 days ago

    Permissions

    No special permissions declared or detected

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    The small generalist — a 384-dimension model. It is fast, cheap, and good enough to reveal whether your pipeline works. Upgrade only when measured recall says the small model is the bottleneck.

    Only when your domain vocabulary is genuinely unusual, you have a few hundred labeled pairs, and the base model demonstrably fails on your queries. Otherwise ship the base model.

    Dense embeddings miss exact matches — product codes, names, IDs. Most production systems end up hybrid: dense retrieval plus keyword search covering each other's failure modes.

    New in LLM Ops