feat: configurable clip length in UI

src-tauri/src/lib.rs

@@ -432,6 +432,7 @@ async fn generate_candidates(
     api_key: Option<String>,
     provider: Option<String>,
     model_name: Option<String>,
+    target_seconds: Option<f64>,
     _allow_demo: bool,
 ) -> Result<Vec<Candidate>, String> {
     let db = state.db.clone();
@@ -442,6 +443,11 @@ async fn generate_candidates(
     let normalized: NormalizedTranscript =
         serde_json::from_str(&transcript.raw_json).map_err(to_command_error)?;
 
+    // Desired clip length from the UI, with a ±25% tolerance band for the model.
+    let target = target_seconds.unwrap_or(60.0).clamp(10.0, 600.0);
+    let min_seconds = (target * 0.75).max(5.0);
+    let max_seconds = target * 1.25;
+
     let active_provider = provider
         .or_else(|| std::env::var("LLM_PROVIDER").ok())
         .unwrap_or_else(|| "deepseek".to_string())
@@ -452,23 +458,23 @@ async fn generate_candidates(
             let key = api_key
                 .or_else(|| std::env::var("ANTHROPIC_API_KEY").ok())
                 .ok_or_else(|| "Set ANTHROPIC_API_KEY or supply Claude API Key to generate candidates.".to_string())?;
-            llm::detect_candidates_with_claude(&normalized, &key)
+            llm::detect_candidates_with_claude(&normalized, &key, min_seconds, max_seconds)
                 .await
                 .map_err(to_command_error)?
         }
         "local" | "ollama" => {
             let model = model_name
                 .or_else(|| std::env::var("OLLAMA_MODEL").ok())
                 .unwrap_or_else(|| "llama3.2".to_string());
-            llm::detect_candidates_with_local_llm(&normalized, &model)
+            llm::detect_candidates_with_local_llm(&normalized, &model, min_seconds, max_seconds)
                 .await
                 .map_err(to_command_error)?
         }
         _ => {
             let key = api_key
                 .or_else(|| std::env::var("DEEPSEEK_API_KEY").ok())
                 .ok_or_else(|| "Set DEEPSEEK_API_KEY or supply DeepSeek API Key to generate candidates.".to_string())?;
-            llm::detect_candidates_with_deepseek(&normalized, &key)
+            llm::detect_candidates_with_deepseek(&normalized, &key, min_seconds, max_seconds)
                 .await
                 .map_err(to_command_error)?
         }
@@ -874,7 +880,10 @@ fn build_drawtext_filters(
         let mut font_option = String::new();
         for path in &font_paths {
             if std::path::Path::new(path).exists() {
-                font_option = format!("fontfile='{}':", path);
+                // Escape the colon in Windows drive-letter paths (e.g. C:/...).
+                // Single quotes protect filtergraph-level separators (commas) but NOT
+                // drawtext's own option separator (':'), so the colon must be backslash-escaped.
+                font_option = format!("fontfile='{}':", path.replace(':', "\\:"));
                 break;
             }
         }

src-tauri/src/llm.rs

@@ -34,12 +34,15 @@ struct DeepseekResponse {
 pub async fn detect_candidates_with_deepseek(
     transcript: &NormalizedTranscript,
     api_key: &str,
+    min_seconds: f64,
+    max_seconds: f64,
 ) -> Result<Vec<CandidateDraft>> {
     let segments = compact_segments(&transcript.segments);
+    let dur = format!("{:.0}-{:.0}", min_seconds, max_seconds);
     let prompt = format!(
         "You are identifying the most viral moments and strongest short-form clip candidates from a long-form transcript. \
 For each candidate, the clip must be self-contained, starting with an extremely engaging hook within the first 3 seconds (to capture immediate attention on social feeds), \
-30-90 seconds long, and cut at clean sentence/thought boundaries. Favor highly shareable content: concrete stories, \
+{dur} seconds long, and cut at clean sentence/thought boundaries. Favor highly shareable content: concrete stories, \
 strong opinions, emotional turns, surprising or counter-intuitive claims, clear payoffs, and high-energy/dramatic peaks. \
 Avoid rambling setup, context-dependent references, and pure filler. Return up to 10 candidates as JSON matching this schema: \
 {{\"candidates\":[{{\"start\":0.0,\"end\":0.0,\"score\":0.0,\"hook\":\"...\",\"rationale\":\"...\"}}]}}\n\nTranscript:\n{segments}"
@@ -78,11 +81,7 @@ Avoid rambling setup, context-dependent references, and pure filler. Return up t
         .map(|c| c.message.content.clone())
         .ok_or_else(|| anyhow!("DeepSeek response did not include choices content"))?;
 
-    let min_duration = if transcript.duration < 60.0 {
-        (transcript.duration * 0.5).max(5.0)
-    } else {
-        30.0
-    };
+    let min_duration = min_clip_duration(transcript.duration, min_seconds);
     parse_candidate_json(&text, min_duration)
 }
 
@@ -95,12 +94,15 @@ struct ClaudeMessage<'a> {
 pub async fn detect_candidates_with_claude(
     transcript: &NormalizedTranscript,
     api_key: &str,
+    min_seconds: f64,
+    max_seconds: f64,
 ) -> Result<Vec<CandidateDraft>> {
     let segments = compact_segments(&transcript.segments);
+    let dur = format!("{:.0}-{:.0}", min_seconds, max_seconds);
     let prompt = format!(
         "You are identifying the most viral moments and strongest short-form clip candidates from a long-form transcript. \
 For each candidate, the clip must be self-contained, starting with an extremely engaging hook within the first 3 seconds (to capture immediate attention on social feeds), \
-30-90 seconds long, and cut at clean sentence/thought boundaries. Favor highly shareable content: concrete stories, \
+{dur} seconds long, and cut at clean sentence/thought boundaries. Favor highly shareable content: concrete stories, \
 strong opinions, emotional turns, surprising or counter-intuitive claims, clear payoffs, and high-energy/dramatic peaks. \
 Avoid rambling setup, context-dependent references, and pure filler. Return up to 10 candidates as JSON only: \
 {{\"candidates\":[{{\"start\":0,\"end\":0,\"score\":0.0,\"hook\":\"...\",\"rationale\":\"...\"}}]}}\n\nTranscript:\n{segments}"
@@ -141,11 +143,7 @@ Avoid rambling setup, context-dependent references, and pure filler. Return up t
         .find_map(|content| content.text)
         .ok_or_else(|| anyhow!("Claude response did not include text content"))?;
 
-    let min_duration = if transcript.duration < 60.0 {
-        (transcript.duration * 0.5).max(5.0)
-    } else {
-        30.0
-    };
+    let min_duration = min_clip_duration(transcript.duration, min_seconds);
     parse_candidate_json(&text, min_duration)
 }
 
@@ -162,18 +160,20 @@ struct OllamaResponse {
 pub async fn detect_candidates_with_local_llm(
     transcript: &NormalizedTranscript,
     model_name: &str,
+    min_seconds: f64,
+    max_seconds: f64,
 ) -> Result<Vec<CandidateDraft>> {
     let segments = compact_segments(&transcript.segments);
-    
-    let system_instructions = "You are identifying the most viral moments and strongest short-form clip candidates from a long-form transcript. \
+
+    let system_instructions = format!("You are identifying the most viral moments and strongest short-form clip candidates from a long-form transcript. \
 For each candidate, the clip must be self-contained, starting with an extremely engaging hook within the first 3 seconds (to capture immediate attention on social feeds), \
-30-90 seconds long, and cut at clean sentence/thought boundaries. \
-CRITICAL: Each clip candidate MUST have a duration between 30 and 90 seconds (i.e. 'end' minus 'start' must be between 30.0 and 90.0). \
-Do NOT return short clips of less than 30 seconds. Combine multiple adjacent sentences to build a meaningful segment of 30-90 seconds. \
+{min_seconds:.0}-{max_seconds:.0} seconds long, and cut at clean sentence/thought boundaries. \
+CRITICAL: Each clip candidate MUST have a duration between {min_seconds:.0} and {max_seconds:.0} seconds (i.e. 'end' minus 'start' must be between {min_seconds:.1} and {max_seconds:.1}). \
+Do NOT return short clips of less than {min_seconds:.0} seconds. Combine multiple adjacent sentences to build a meaningful segment of {min_seconds:.0}-{max_seconds:.0} seconds. \
 Favor highly shareable content: concrete stories, strong opinions, emotional turns, surprising or counter-intuitive claims, clear payoffs, and high-energy/dramatic peaks. \
 Avoid rambling setup, context-dependent references, and pure filler. \
 You MUST identify and return at least 3-5 candidates (up to 10 candidates). Do not return an empty candidates list. \
-Ensure the 'start' and 'end' values correspond to actual timestamps in the transcript. Do not output 0.0 for start and end times.";
+Ensure the 'start' and 'end' values correspond to actual timestamps in the transcript. Do not output 0.0 for start and end times.");
 
     let user_content = format!("Transcript:\n{}", segments);
 
@@ -227,14 +227,21 @@ Ensure the 'start' and 'end' values correspond to actual timestamps in the trans
     }
 
     let res_body: OllamaResponse = response.json().await.context("parsing local Ollama response")?;
-    let min_duration = if transcript.duration < 60.0 {
-        (transcript.duration * 0.5).max(5.0)
-    } else {
-        30.0
-    };
+    let min_duration = min_clip_duration(transcript.duration, min_seconds);
     parse_candidate_json(&res_body.message.content, min_duration)
 }
 
+/// Minimum acceptable clip length, relaxed for transcripts shorter than the
+/// requested band so very short sources still yield candidates.
+// ponytail: simple heuristic; revisit if users want a hard floor regardless of source length
+fn min_clip_duration(transcript_duration: f64, min_seconds: f64) -> f64 {
+    if transcript_duration < min_seconds * 2.0 {
+        (transcript_duration * 0.5).max(5.0)
+    } else {
+        min_seconds
+    }
+}
+
 pub fn demo_candidates(transcript: &NormalizedTranscript) -> Vec<CandidateDraft> {
     transcript
         .segments

src/main.tsx

@@ -131,6 +131,9 @@ function App() {
   const [localLlmModel, setLocalLlmModel] = useState(() => {
     return localStorage.getItem("autoshorts_local_llm_model") || "llama3.2";
   });
+  const [clipLength, setClipLength] = useState(() => {
+    return Number(localStorage.getItem("autoshorts_clip_length")) || 60;
+  });
   const [deepgramKey, setDeepgramKey] = useState(() => {
     return localStorage.getItem("autoshorts_deepgram_key") || "";
   });
@@ -201,6 +204,10 @@ function App() {
     localStorage.setItem("autoshorts_local_llm_model", localLlmModel);
   }, [localLlmModel]);
 
+  useEffect(() => {
+    localStorage.setItem("autoshorts_clip_length", String(clipLength));
+  }, [clipLength]);
+
   useEffect(() => {
     localStorage.setItem("autoshorts_deepgram_key", deepgramKey);
   }, [deepgramKey]);
@@ -365,6 +372,7 @@ function App() {
         apiKey: activeKey || null,
         provider: llmEngine,
         modelName: llmEngine === "local" ? localLlmModel.trim() : null,
+        targetSeconds: clipLength,
         allowDemo: false,
       });
       await refresh(projectId);
@@ -447,6 +455,7 @@ function App() {
           apiKey: activeKey || null,
           provider: llmEngine,
           modelName: llmEngine === "local" ? localLlmModel.trim() : null,
+          targetSeconds: clipLength,
           allowDemo,
         });
         await refresh(detail.project.id);
@@ -634,6 +643,17 @@ function App() {
                         <option value="deepseek">DeepSeek (Cloud)</option>
                       </select>
                     </label>
+                    <label>
+                      <span>Clip Length (seconds)</span>
+                      <input
+                        value={clipLength}
+                        onChange={(event) => setClipLength(Number(event.target.value) || 0)}
+                        onBlur={() => setClipLength((v) => Math.min(600, Math.max(10, v || 60)))}
+                        type="number"
+                        min={10}
+                        max={600}
+                      />
+                    </label>
                     {transcriptionEngine === "deepgram" && (
                       <label>
                         <span>Deepgram API Key</span>