Skip to content
ollama-intern-mcp

ollama_corpus_answer

ollama_corpus_answer is the synthesis flagship. Give it a corpus name and a question; it retrieves the top-k chunks, then synthesizes an answer grounded only in those chunks. Every claim cites a chunk id. Outside knowledge is forbidden by the system prompt and the citations are validated server-side — invalid ids are stripped before the result returns.

Tier — Deep + Embed

Section titled “Tier — Deep + Embed”

Embed tier for retrieval (nomic-embed-text on the default profile), Deep tier for synthesis (hermes3:8b, larger context). Two model loads back-to-back.

When to use it

Section titled “When to use it”
  • “What does our docs say about X?” — over a pre-indexed corpus
  • “Summarize what we’ve decided about Y” — over project memory
  • “Give me a chunk-grounded answer about Z” — when you need citation trail

When NOT to use it

Section titled “When NOT to use it”
  • You have specific files in mind → ollama_research (path-grounded)
  • The corpus isn’t indexed yet → run ollama_corpus_index first
  • You need a one-shot answer with no grounding → ollama_chat (last resort)

Schema

Section titled “Schema”
{
  corpus:      string;        // name; matches /^[a-zA-Z0-9_-]+$/
  question:    string;        // max 1000 chars
  mode?:       "hybrid" | "vector" | "lexical";  // retrieval strategy
  top_k?:      number;        // 1..20, default 5
  max_words?:  number;        // 20..1000, default 200 (target answer length)
  min_top_score?: number;     // 0..1, topicality floor; v2.2.0+
  model?:      string;        // per-call model override (advanced)
}

Full source: src/tools/corpusAnswer.ts.

Example call

Section titled “Example call”
{
  "tool": "ollama_corpus_answer",
  "arguments": {
    "corpus": "ollama-docs",
    "question": "How is residency reported in /api/ps?",
    "top_k": 6,
    "min_top_score": 0.5
  }
}

Envelope result:

{
  "answer": "/api/ps reports loaded models with size_vram, size, and digest...",
  "citations": [
    {
      "chunk_id":     "abc123",
      "source_path":  "/repo/docs/api/ps.md",
      "heading_path": ["API", "/api/ps"],
      "score":        0.84
    },
    {
      "chunk_id":     "def456",
      "source_path":  "/repo/docs/concepts/residency.md",
      "score":        0.71
    }
  ],
  "weak":      false,
  "abstained": false,
  "retrieval": { "mode": "hybrid", "top_k": 6, "hits_returned": 6 }
}

Abstention contract (v2.2.0+)

Section titled “Abstention contract (v2.2.0+)”

The tool can return three honest non-answers:

ShapeWhenWhat you should do
weak: trueFewer than 2 retrieval hitsTreat the answer as a hint, not authoritative
abstained: true, min_top_score exceededTop hit fell below the threshold you setLower the threshold, expand the corpus, or accept the abstention
abstained: true, zero hitsNo chunks matched at allIndex more relevant content into the corpus

abstained: true always comes with an empty answer and empty citations[]. It is success, not failure — the tool refusing to launder weak retrieval into authoritative output.

Common pitfalls

Section titled “Common pitfalls”

abstained: true with min_top_score you don’t remember setting. The default is no floor; abstention here means the corpus genuinely didn’t address the question. Either expand the corpus (index more files) or relax min_top_score if you set one.

citations[] shorter than the answer’s claim count. Some claims got synthesized from the same chunk; that’s normal. If you see zero citations with a non-empty answer, that’s the abstention contract — the model declined.

weak: true on an answer that looks fine. Two-or-fewer chunks matched the query. The answer might still be correct, but you’re working off a narrow base — spot-check before promoting.

Corpus name has invalid characters. Names must match [a-zA-Z0-9_-]+. Spaces, slashes, dots all rejected at the schema level.

question longer than 1000 chars. Capped to keep the answer-from- corpus contract intact. Beyond 1000 chars you’re asking the model to summarize the question itself, which defeats the grounding.

Section titled “Related tools”