Skip to content
ollama-intern-mcp

ollama_classify

ollama_classify is the Instant-tier label-picker. Give it text and a list of candidate labels; it returns the best match plus a confidence score.

Tier — Instant

Section titled “Tier — Instant”

Fast. hermes3:8b on the default profile, num_ctx: 4096 for the dev-rtx5080 profile. Designed for high-throughput batch work.

When to use it

Section titled “When to use it”
  • Routing tickets / logs / messages into a fixed taxonomy
  • Spam / quality / sentiment-style binary or N-way classification
  • Pre-filtering inputs before more expensive tools run

When NOT to use it

Section titled “When NOT to use it”
  • The labels are open-ended → ollama_extract with a string field
  • You need extracted reasoning, not just a label → ollama_extract
  • The text is a long file → use source_path mode (server reads it for you)

Schema

Section titled “Schema”
{
  // Exactly one of {text, source_path, items} must be provided.
  text?:        string;       // single text, classified directly
  source_path?: string;       // server reads + classifies the file
  items?: Array<{
    id: string;               // caller-provided, unique within the batch
    text: string;
  }>;                          // batch mode


  labels:       string[];     // min 2; the model picks exactly one (or null)
  allow_none?:  boolean;      // if true, return label=null when conf < threshold
  threshold?:   number;       // 0..1; default 0.7
  frame?:       string;       // v2.2.0+ — what the classification is FOR; off-topic → null
  per_file_max_chars?: number; // bytes to read for source_path (default 40k)
  model?:       string;       // per-call model override (advanced)
}

Full source: src/tools/classify.ts.

Example call (single)

Section titled “Example call (single)”
{
  "tool": "ollama_classify",
  "arguments": {
    "text": "the build failed because the API returned 502",
    "labels": ["infra", "code", "user-error"]
  }
}

Envelope result:

{
  "label": "infra",
  "confidence": 0.93,
  "allow_none": false,
  "threshold": 0.7
}

Example call (batch)

Section titled “Example call (batch)”
{
  "tool": "ollama_classify",
  "arguments": {
    "labels": ["bug", "feature-request", "question", "spam"],
    "items": [
      { "id": "t-001", "text": "login button does nothing on mobile Safari" },
      { "id": "t-002", "text": "can you add dark mode?" },
      { "id": "t-003", "text": "BUY NOW AMAZING CRYPTO DEAL" }
    ]
  }
}

Envelope result:

{
  "items": [
    { "id": "t-001", "ok": true, "result": { "label": "bug",             "confidence": 0.94 } },
    { "id": "t-002", "ok": true, "result": { "label": "feature-request", "confidence": 0.91 } },
    { "id": "t-003", "ok": true, "result": { "label": "spam",            "confidence": 0.97 } }
  ]
}

Example call (frame-bound, v2.2.0+)

Section titled “Example call (frame-bound, v2.2.0+)”

When frame is supplied, the model first decides whether the text is on-topic for the frame, and returns null (with off_topic: true) if it’s not — even if a label would technically fit.

{
  "tool": "ollama_classify",
  "arguments": {
    "text": "this PR refactors the date-formatting utility",
    "labels": ["security-relevant", "not-security-relevant"],
    "frame": "Should this PR get a security review?"
  }
}

Returns { label: "not-security-relevant", confidence: 0.88, frame_alignment: { on_topic: true } }.

Common pitfalls

Section titled “Common pitfalls”

SCHEMA_INVALID on duplicate ids in items[]. Batch inputs require unique caller-provided ids. Pick stable ids (e.g. row IDs from your DB). There is no implicit dedup.

Confidence below threshold returning a label anyway. Set allow_none: true to opt into the null-return-when-uncertain behavior. Without it, the threshold is informational only.

Using text for a 50K-char document. Use source_path instead — the server reads + truncates per per_file_max_chars (default 40k chars). You save Claude context AND avoid the prompt-too-long error.

Both text and items provided. Returns SCHEMA_INVALID. Pick one input mode per call.

Section titled “Related tools”