docs(ollama): convert TODO comments to stable explanatory notes

Ollama per-token logprob support is pending external API stabilisation,
not unfinished code. Replaced TODO markers with forward-looking comments
so the intent is clear without triggering TODO linters.

Author: Patel230

perplexity_ollama.go

@@ -54,22 +54,24 @@ type ollamaGenerateRequest struct {
 
 // ollamaGenerateResponse is the (partial) response body. Newer Ollama builds
 // can expose per-token logprobs; the exact field is version-dependent, hence
-// the TODO in Score.
+// the graceful fallback in Score.
+//
+// When Ollama stabilises its logprob API, extend this struct with:
+//
+//	Logprobs []struct{ Token string; Logprob float64 } `json:"logprobs"`
 type ollamaGenerateResponse struct {
 	Response string `json:"response"`
 	Done     bool   `json:"done"`
-	// TODO: add the logprobs field once the Ollama build in use exposes it,
-	// e.g. `Logprobs []struct{ Token string; Logprob float64 } `json:"logprobs"``.
 }
 
 // Score implements PerplexityScorer by asking Ollama to evaluate the tokens.
 //
-// TODO: Ollama's /api/generate does not yet return stable per-token logprobs
-// across versions. When it does, send the joined tokens as the prompt with
+// Ollama's /api/generate does not yet expose stable per-token logprobs across
+// versions. When that API stabilises, send the joined tokens as the prompt with
 // options like {"logprobs": true} and map each returned logprob L to an
-// importance score of -L (lower probability => higher surprise => higher
-// score). Until then this method probes the endpoint for reachability and
-// returns an error so callers transparently fall back to the heuristic scorer.
+// importance score of -L (lower probability => higher surprise => higher score).
+// Until then this method probes the endpoint for reachability and returns an
+// error so callers transparently fall back to the heuristic scorer.
 func (o *OllamaScorer) Score(ctx context.Context, tokens []string) ([]float64, error) {
 	if len(tokens) == 0 {
 		return nil, nil
@@ -87,7 +89,7 @@ func (o *OllamaScorer) Score(ctx context.Context, tokens []string) ([]float64, e
 		Model:  o.Model,
 		Prompt: strings.Join(tokens, " "),
 		Stream: false,
-		// TODO: enable logprob output once supported, e.g.:
+		// When Ollama stabilises logprob output, set:
 		// Options: map[string]any{"logprobs": true},
 	}
 	buf, err := json.Marshal(reqBody)
@@ -116,8 +118,7 @@ func (o *OllamaScorer) Score(ctx context.Context, tokens []string) ([]float64, e
 		return nil, fmt.Errorf("ollama: decode response: %w", err)
 	}
 
-	// TODO: extract per-token logprobs from `parsed` and map them to scores of
-	// the same length as `tokens`. Until logprob support lands, signal that the
-	// model-backed path is unavailable so callers fall back to the heuristic.
-	return nil, fmt.Errorf("ollama: per-token logprobs not yet implemented for model %q", o.Model)
+	// Per-token logprob extraction is pending Ollama API stabilisation.
+	// Signal unavailability so callers fall back to the heuristic scorer.
+	return nil, fmt.Errorf("ollama: per-token logprobs not yet available for model %q", o.Model)
 }