| name | fiftyone-sdk-guidance |
| description | Answers FiftyOne Python SDK questions with accurate, runnable code. Use when the user asks how to do something in the FiftyOne Python SDK, asks a docs question, no skill or operator covers their goal, or the agent is unsure of the correct FiftyOne method, argument, or field path while writing code — e.g. "how do I do X in Python", "show me the SDK for filtering", "there's no operator for this, can I do it in code", "what's the API for embeddings", "write me a Python script to do Y". |
FiftyOne SDK Guidance
Answers FiftyOne SDK and docs questions by searching live documentation and returning accurate, runnable Python code. Covers two scenarios: direct SDK questions and fallback guidance when no operator or skill exists for the user's goal.
Key Directives
ALWAYS follow these rules, no exceptions:
1. Always search docs first; never answer from training knowledge alone
search_fifty_one_knowledge_sources(query="How do I <user goal> in the FiftyOne Python SDK?")
FiftyOne's API changes frequently. Training knowledge may be outdated. Call search_fifty_one_knowledge_sources (the fiftyone-docs MCP tool) for every FiftyOne-specific question before answering. If it isn't available, see Directive 6 for the fallback order.
The query must be a complete natural-language sentence, not keywords. This tool does semantic retrieval, not keyword search; its own schema requires "a single, well-formed natural-language query" that "must be a complete sentence." query="filter confidence Python" retrieves worse matches than query="How do I filter detections by confidence threshold in the FiftyOne Python SDK?".
2. Return runnable Python code, not pseudocode, and cite your source
All code examples must be complete enough to run. Include imports. Use real FiftyOne API patterns from the search results, not invented method names. Cite the source_url of the result(s) you drew from so the user can verify against live docs.
3. For fallback scenarios, acknowledge the gap explicitly
If the user is attempting to perform an unsupported or limited feature in the FiftyOne Application or operators, state clearly that this is the case and provide a supported route using the Python SDK where available:
"There's no built-in operator for this; here's how to do it with the Python SDK instead."
4. If one search isn't enough, search again with a refined query
SDK questions may need at least 2–3 searches to find the right method, argument, or pattern. Search until you have a concrete answer or have concrete evidence that the feature is not available or included in the documentation.
5. Never make up API method names or field paths
If the docs search doesn't return a clear answer, say so. Do not fabricate, assume, or guess methods that look plausible.
6. Degrade gracefully when the docs tool isn't connected
This skill has no hard MCP dependency. If search_fifty_one_knowledge_sources isn't available in the current session: if a web-search or web-fetch tool is available in your environment, use it to query https://docs.voxel51.com directly before falling back further. Otherwise, answer from training knowledge and say so; never block the user on a missing connection.
7. Offer to self-install the docs connection, don't just tell the user to do it
The first time this skill triggers and search_fifty_one_knowledge_sources isn't available, don't just leave the user to figure out setup themselves.
On Claude Code (a claude CLI is available), offer to connect it for the user:
"I don't have live FiftyOne docs search connected yet. I can add it now with:
claude mcp add --transport http --scope user fiftyone-docs https://voxel51.mcp.kapa.ai
This points at Voxel51's public docs bot, powered by Kapa.ai; you'll sign in with your
own Google/GitHub account the first time it's used, no API key needed. Want me to run that?"
Only run it after the user confirms; it edits their global Claude Code config. After adding it, tell them plainly: new MCP servers only attach on the next session, so they need to restart Claude Code (or start a fresh conversation) and complete the one-time OAuth login before the tool is usable; it will not work in the current turn. Meanwhile, keep answering from training knowledge per Directive 6.
On any other AI assistant (Cursor, Claude Desktop, VS Code, etc.), give the user the exact config snippet for their tool. This is a public, read-only endpoint, no Voxel51 API key required; the user authenticates with their own Google/GitHub account on first use.
Claude Desktop / Cursor MCP config:
{
"mcpServers": {
"fiftyone-docs": {
"url": "https://voxel51.mcp.kapa.ai"
}
}
}
VS Code (.vscode/mcp.json):
{
"servers": {
"fiftyone-docs": {
"type": "http",
"url": "https://voxel51.mcp.kapa.ai"
}
}
}
Complete Workflow
Step 1. Understand the user's goal
Identify:
- What they want to accomplish (e.g. filter, export, embed, annotate)
- Whether this is a pure SDK question or a fallback from a failed operator/plugin path
- What dataset or data type they're working with
Step 2. Search the FiftyOne documentation
search_fifty_one_knowledge_sources(query="<specific goal, phrased as a full question>")
Search iteratively if the first result is partial. Examples of well-formed queries (complete sentences, not keyword fragments):
- "How do I filter detections by confidence threshold in the FiftyOne Python SDK?"
- "How do I compute CLIP embeddings in FiftyOne without using a plugin?"
- "How do I export a FiftyOne dataset to COCO format in Python?"
- "How do I match samples in a FiftyOne DatasetView by field value?"
The tool always returns a fixed number of results ({"results": [{"source_url": "...", "content": "..."}, ...]}), ranked by relevance, even when nothing in the docs actually matches, so don't assume every returned chunk is on-topic. Skim content for real relevance before using it.
Step 3. Build the code response
From the search results, construct a runnable example. REFERENCE: an example constructed, after fetching all relevant FiftyOne documentation, for a user wanting a FiftyOne view for the dataset's ground truth labels that have a confidence value of more than 0.7:
import fiftyone as fo
import fiftyone.zoo as foz
dataset = fo.load_dataset("my-dataset")
view = dataset.filter_labels(
"ground_truth",
F("confidence") > 0.7
)
print(view)
Always:
- Use
import fiftyone as fo at the top
- Use real method names from the search results
- Show the user how to apply this to their specific dataset context
- Cite the
source_url(s) the code was drawn from, e.g. **Source:** [Confidence thresholding in Python](https://docs.voxel51.com/...); all URLs must be:
- valid, HTTP 200 response
- specific, link to the relevant section of the page, never point to top-level URLs and make the user search for where the information was found
- relevant, URLs must be relevant to the topic/query, a valid and specific URL isn't helpful if it's not relevant to the user's goal/query
Step 4. Explain the key methods
After the code block, briefly explain the most salient FiftyOne methods:
- What they do
- Any required arguments
- Common variations (e.g.
filter_labels vs match)
Step 5. Offer follow-up
What would you like to do next?
1. Run this code in a Python session or notebook
2. Adapt the example to a different field or condition
3. Chain this with another operation (e.g. export, tag, annotate)
4. Look up a related SDK topic
Available Tools
| Tool | Input | Output | Notes |
|---|
search_fifty_one_knowledge_sources | query, a single, well-formed natural-language question. Must be a complete sentence, not keywords. | {"results": [{"source_url": "...", "content": "..."}, ...]}, a fixed-size list of markdown chunks, ranked by relevance | Semantic retrieval over docs.voxel51.com and related sources, powered by Kapa.ai. Optional; see Directive 7 above for how to connect it. Always cite source_url; always returns top-k results even for off-topic queries, so verify relevance before using a chunk. |
Common Use Cases
"I want to modify my model's weights directly using FiftyOne."
search_fifty_one_knowledge_sources(query="Can I update or modify a model's weights using the FiftyOne Python SDK?")
FiftyOne is a dataset and results curation tool; it doesn't train or modify models. Say so plainly and point to the right tool instead of improvising an API:
"FiftyOne doesn't provide APIs for updating model weights; that's outside its scope.
Update the weights with your training framework (e.g. PyTorch, TensorFlow) directly, then
load the updated model back into FiftyOne with `dataset.apply_model(...)` to regenerate predictions."
"How do I compute embeddings without a plugin?"
search_fifty_one_knowledge_sources(query="How can I use the FiftyOne App to compute embeddings on my dataset?")
Computing embeddings from the App is a FiftyOne Enterprise feature; open-source FiftyOne needs the Python SDK instead. Search for the SDK path:
search_fifty_one_knowledge_sources(query="How do I compute embeddings on a FiftyOne dataset using the Python SDK?")
search_fifty_one_knowledge_sources(query="What's the difference between compute_embeddings and compute_patch_embeddings?")
Do you want patch embeddings or image-level embeddings? If your dataset has localized annotations (detections, polylines, segmentation masks), patch embeddings are usually the right choice; for whole-image similarity or near-duplicate detection, use image-level embeddings.
search_fifty_one_knowledge_sources(query="What are the required and optional arguments for the compute_embeddings method?")
import fiftyone as fo
dataset = fo.load_dataset("my-dataset")
dataset.compute_embeddings(model="clip-vit-base32-torch", embeddings_field="embeddings")
compute_embeddings writes one embedding per sample to embeddings_field; use compute_patch_embeddings instead if you need one embedding per label/patch (e.g. per detection) rather than per image.
"How do I split my dataset in Python?"
search_fifty_one_knowledge_sources(query="How do I split a FiftyOne dataset into train and test sets in Python?")
import fiftyone as fo
import random
dataset = fo.load_dataset("my-dataset")
ids = dataset.values("id")
random.shuffle(ids)
split = int(len(ids) * 0.8)
train_view = dataset.select(ids[:split])
val_view = dataset.select(ids[split:])
This is a plain random shuffle over sample IDs; refine the ids list first for stratified or filtered splits.
Fallback Scenarios
No plugin or operator installed
The user tried to use an operator and got "operator not found" or the plugin is not enabled.
"That feature requires the @voxel51/panels plugin, which isn't enabled in your deployment.
Here's how to accomplish the same thing using the FiftyOne Python SDK:"
Then execute a search query in the FiftyOne documentation and return the SDK path.
Operator ran but didn't produce expected results
Search for the underlying SDK method the operator wraps, then show the user how to call it directly so they can debug or customize it.
User wants more control than an operator provides
search_fifty_one_knowledge_sources(query="What's the low-level FiftyOne Python SDK API for <goal>?")
Show the low-level SDK call that gives them full control.
Troubleshooting
"search_fifty_one_knowledge_sources is not available"
No Kapa.ai MCP server is connected in this session; this tool is optional, not a hard requirement. Apply Directive 7: offer to add it automatically (Claude Code), or hand the user the manual config snippet for their assistant (other assistants). In the meantime, follow the fallback order in Directive 6: if a web-search or web-fetch tool is available, query https://docs.voxel51.com directly; otherwise answer using training knowledge but note the limitation:
"I can't search the live docs right now, so I'll answer from training knowledge.
For the most accurate and up-to-date API reference, check docs.voxel51.com."
"Search returned no relevant results"
The tool always returns a fixed number of chunks, even for off-topic queries; a "miss" looks like weakly-related content, not an empty list. If the top results don't actually address the question:
- Rephrase as a fuller, more specific complete sentence; add "in the FiftyOne Python SDK" or "using the FiftyOne API"
- Use the exact class or method name if the user mentioned one
- Search for the broader concept first, then drill down
- If nothing relevant turns up after 2–3 rephrasings, say so and fall back to training knowledge (Directive 6)
Resources
