Code search

Nimble code search helps a code search tool decide what to show a coding agent. You send a query and a list of items, where each item is a piece of a file or a folder. For each item, Nimble returns the probability that the agent needs it for the query. Send the request to POST /v1/systemone with one of the code search models. The request uses the System One format with noul questions, so clients that already send System One requests can use it.

Make a request

Question q0 is about item n0, question q1 is about item n1, and so on.

curl https://api.bespokelabs.ai/v1/systemone \
  -H "Authorization: Bearer $BESPOKE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nimble-codegrep",
    "state": {
      "query": "Uploads retry forever after a 413 response.",
      "items": [
        {
          "id": "n0",
          "path": "src/upload/retry.py",
          "kind": "file",
          "filePreview": {"text": "def backoff(attempt):\n    while True:\n        retry()"}
        },
        {
          "id": "n1",
          "path": "src/upload",
          "kind": "directory",
          "childPreview": {"entries": [{"name": "retry.py", "kind": "file"}]}
        }
      ]
    },
    "questions": {
      "q0": {"type": "noul", "instructions": "Does the file src/upload/retry.py contain code needed for this query?"},
      "q1": {"type": "noul", "instructions": "Is the directory src/upload worth exploring for this query?"}
    }
  }'

from bespokelabs import BespokeLabs

client = BespokeLabs()  # reads BESPOKE_API_KEY

response = client.nimble.system_one(
    model="nimble-codegrep",
    state={
        "query": "Uploads retry forever after a 413 response.",
        "items": [
            {
                "id": "n0",
                "path": "src/upload/retry.py",
                "kind": "file",
                "filePreview": {"text": "def backoff(attempt):\n    while True:\n        retry()"},
            },
            {
                "id": "n1",
                "path": "src/upload",
                "kind": "directory",
                "childPreview": {"entries": [{"name": "retry.py", "kind": "file"}]},
            },
        ],
    },
    questions={
        "q0": {"type": "noul", "instructions": "Does the file src/upload/retry.py contain code needed for this query?"},
        "q1": {"type": "noul", "instructions": "Is the directory src/upload worth exploring for this query?"},
    },
)
keep = [qid for qid, answer in response.answers.items() if answer.noul > 0.5]
print(keep)

Read the answer

{
  "model": "nimble-codegrep",
  "effort": "medium",
  "answers": {
    "q0": { "type": "noul", "noul": 0.99 },
    "q1": { "type": "noul", "noul": 0.99 }
  },
  "details": {
    "q0": { "raw": 0.85, "overflow": false, "escalated": false, "scores": { "nimble-codegrep-4b": 0.85 } },
    "q1": { "raw": 0.82, "overflow": false, "escalated": false, "scores": { "nimble-codegrep-4b": 0.82 } }
  },
  "usage": { "input_tokens": 472, "output_tokens": 2 },
  "request_id": "0789d4c7-399f-4e70-9721-9f21bab043bc",
  "escalation_skipped": false
}

Keep the items whose noul is above 0.5. Nimble sets the scores so that this keeps about 95% of the items that an agent needs.
details.raw is the model’s own probability before that adjustment.
details.escalated says whether a larger model scored the item.
An item whose prompt is longer than 8,192 tokens is not scored. Its noul is 1.0, so you keep it, and overflow is true.

With the Python SDK, the extra fields are attributes of the response, e.g., response.details.

Models

Model	How it answers
`nimble-codegrep-lite`	A 4B model scores every item. This is the fastest and the cheapest.
`nimble-codegrep`	The 4B model scores every item. Then a 27B model scores again the items that the 4B model is unsure about.
`nimble-codegrep-max`	The 27B model scores every item. This is the most accurate.

The 27B model is running on weekdays from 07:30 to 19:00 Pacific time. At other times, it must start before it can answer. Until it is ready, a nimble-codegrep-max request gets 503 with a Retry-After header. So does a nimble-codegrep request with an item that the 4B model is unsure about. Starting takes a few minutes. You are not charged for a 503.

Limits

Limit	Value
Questions in a request	1 to 128
Characters in the state	2,000,000
Request body	1 MiB
Prompt tokens for one item	8,192

Use noul questions. Code search also accepts boolean questions, and it answers them the same way, with a noul answer. Send the state as an object with a query and items. Every item needs an id, a path and a kind. Question q<i> is about item n<i>, and a question ID with no matching item gets 422. Code search also accepts a string state and uses it whole for every question, but the models were not trained on that.

Get started

Guides

Make a request

Read the answer

Models

Limits

​Make a request

​Read the answer

​Models

​Limits

Make a request

Read the answer

Models

Limits