More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Chroma Vector Retrieval for RAG

    1

    Build retrieval that actually retrieves: design Chroma collections around metadata-first filtering, chunk documents for recall, tune queries, and measure recall before trusting a single RAG answer.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Chroma Vector Retrieval for RAG

    Chroma Vector Retrieval for RAG

    Example session with this skill installed

    We want document Q&A over our admin guide (versioned, sectioned) and our support ticket archive (short, timestamped). Build the retrieval so answers come from the right chunk.

    • Read your context and instructions
    • Compiled the chroma vector retrieval

    DESIGN: Two document types, two collections. product-docs for the guide: IDs like admin-guide#security#key-rotation, metadata with source, version, section, updated date. support-tickets: IDs with ticket number and timestamp, metadata with status, product area, date. Persistent client from the start — production, not a notebook.

    CHUNKING: Guide chunks follow sections, 15% overlap, section path prepended so each chunk reads alone. Tickets are one chunk each. A first pass used fixed windows; the read-alone test caught chunks ending mid-procedure, so the scheme switched to sections.

    QUERY: User questions filter on version 2.4 first, then rank. Budget 5. Without the filter, the top hit for key rotation was a version 1.9 chunk describing a removed flow — the version trap, caught in testing.

    EVAL: 25 questions with known-good chunks. Recall at 5: 84% first run. Failures classified: three wrong-chunk (two fixed by splitting an overloaded chunk, one by adding a missing section), one stale chunk (version filter gap on a re-ingested page). After fixes, recall at 5 is 92% — above the ship gate. The eval report ships with the collection, including the two remaining failures and their planned fixes.

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

    About this skill

    The embedding model gets the glory, but retrieval quality is decided by unglamorous choices: how you chunk, what metadata you store, and whether you measure. This skill makes those choices in order. Design first: one collection per document type, traceable IDs that map to a source document and a location in it, and metadata stored for every filter you will ever use — because re-embedding a million chunks to add a field is a bad afternoon.

    Chunking follows meaning, not fixed windows: sections, procedures, and Q&A pairs, with 10–20% overlap and a read-alone test that catches chunks ending mid-procedure. Queries filter on metadata before ranking (the version filter alone beats a better embedding model against the stale-chunk trap), retrieve a budgeted 5–10, and normalize embeddings before comparing. Evaluation is mandatory: 20–30 known-answer questions, recall at 5, failures classified as retrieval vs. generation problems, and a ship gate (below 80% blocks shipping). A worked example takes a support Q&A collection from 84% to 92% recall at 5 by fixing chunking and a version-filter gap.

    • Pure nearest-neighbor search at massive scale with no metadata may be faster in a dedicated ANN library.
    • Managed infrastructure with SLAs belongs in a hosted vector database.
    • Embeddings are weak on exact tokens (error codes, part numbers) — those need a keyword filter alongside vector ranking.
    • Retrieval degrades as collections grow; the eval set must be refreshed when docs change.

    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

    Different types have different chunk shapes and metadata. Mixing them means every query fights noise from the wrong type — a support ticket and a product guide never share a chunk shape, so the compromise serves neither.

    Check the eval first. If the right chunk is not in the top 5, it is a retrieval problem — fix chunking and metadata, not the prompt. If the right chunk is there and the answer is still wrong, it is a generation problem.

    Start at 5. Raise it only when the eval shows the right chunk sitting at rank 6–10 for real questions. Retrieving 50 "to be safe" just burns context on noise.

    New in LLM Ops