feat: wip on implementing gemini analyzer

This commit is contained in:
Stevan Freeborn
2025-07-31 00:10:27 -05:00
parent f6773cb64e
commit d7a953f187
10 changed files with 162 additions and 85 deletions
@@ -0,0 +1,88 @@
using System.Text;
using System.Text.Json;
using System.Text.Json.Serialization;
using NAudio.CoreAudioApi;
using StreamShorts.Library.Transcription;
namespace StreamShorts.Library.Analysis;
public sealed class GeminiAnalyzer(
IHttpClientFactory httpClientFactory,
string apiKey
) : ITranscriptAnalyzer
{
private readonly IHttpClientFactory _httpClientFactory = httpClientFactory ?? throw new ArgumentNullException(nameof(httpClientFactory));
private readonly string _apiKey = apiKey ?? throw new ArgumentNullException(nameof(apiKey));
public async Task<TranscriptAnalysis> AnalyzeAsync(IEnumerable<TranscriptionSegment> segments)
{
using var client = _httpClientFactory.CreateClient();
client.Timeout = TimeSpan.FromMinutes(5);
// TODO: Load prompt from resource file
var requestUrl = $"https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-lite:generateContent?key={_apiKey}";
using var request = new HttpRequestMessage(HttpMethod.Post, requestUrl)
{
Content = new StringContent(
JsonSerializer.Serialize(new
{
contents = new[]
{
new
{
role = "user",
parts = new[]
{
new
{
text = prompt
}
},
}
},
generationConfig = new
{
responseMimeType = "application/json",
}
}),
Encoding.UTF8,
"application/json"
)
};
var response = await client.SendAsync(request).ConfigureAwait(false);
var responseContent = await response.Content.ReadAsStringAsync().ConfigureAwait(false);
var responseJson = JsonSerializer.Deserialize<LLMResponse>(responseContent);
var candidatesText = responseJson?
.Candidates?
.FirstOrDefault()?
.Content
.Parts?.FirstOrDefault()?
.Text;
var clips = JsonSerializer.Deserialize<List<ShortClip>>(candidatesText ?? string.Empty);
return new TranscriptAnalysis(clips ?? []);
}
}
// TODO: Sort this shit out
record LLMResponse(
[property: JsonPropertyName("candidates")]
Candidate[] Candidates
);
record Candidate(
[property: JsonPropertyName("content")]
Content Content
);
record Content(
[property: JsonPropertyName("parts")]
Part[] Parts
);
record Part(
[property: JsonPropertyName("text")]
string Text
);
@@ -0,0 +1,8 @@
using StreamShorts.Library.Transcription;
namespace StreamShorts.Library.Analysis;
public interface ITranscriptAnalyzer
{
Task<TranscriptAnalysis> AnalyzeAsync(IEnumerable<TranscriptionSegment> segments);
}
@@ -0,0 +1,16 @@
using System.Text.Json.Serialization;
namespace StreamShorts.Library.Analysis;
public record ShortClip(
[property: JsonPropertyName("title")]
string Title,
[property: JsonPropertyName("description")]
string Description,
[property: JsonPropertyName("explanation")]
string Explanation,
[property: JsonPropertyName("start_time")]
TimeSpan StartTime,
[property: JsonPropertyName("end_time")]
TimeSpan EndTime
);
@@ -0,0 +1,6 @@
namespace StreamShorts.Library.Analysis;
public class TranscriptAnalysis(IEnumerable<ShortClip> shortClips)
{
public IEnumerable<ShortClip> ShortClips { get; init; } = shortClips;
}
@@ -0,0 +1,28 @@
I need your help to transform my YouTube live stream transcript into engaging YouTube Shorts. Act as my content editor and pinpoint **all potential candidate segments** that are perfect for short-form video. I'm looking for clips that are:
  - **Funny:** Moments that will make viewers laugh.
  - **Informative:** Sections packed with valuable information or tips.
  - **Insightful:** Portions offering unique perspectives or 'aha\!' moments.
For each suggested short, please provide:
  - The **start time** of the initial segment and the **end time** of the final segment. The duration of each short should be no longer than 3 minutes, but **aim for durations between 15 seconds and 60 seconds**. However, the short **must be as long as necessary to capture the complete thought or idea**, even if it means exceeding the target range or extending slightly to capture all necessary dialogue.
  - A concise **title** that grabs attention.
  - A brief **description** highlighting the short's content and its appeal.
  - An **explanation** of why this particular segment is suitable for a YouTube Short, focusing on its potential for discoverability and engagement.
Please format your response as a JSON array of objects with the following structure:
```json
{
  "title": "string",
  "start_time": "string",
  "end_time": "string",
  "description": "string",
  "explanation": "string"
}
```
Here is the transcript of my YouTube live stream:
{0}
@@ -1,6 +1,7 @@
<Project Sdk="Microsoft.NET.Sdk">
<ItemGroup>
<PackageReference Include="FFMpegCore" />
<PackageReference Include="Microsoft.Extensions.Http" />
<PackageReference Include="NAudio" />
<PackageReference Include="Whisper.net.AllRuntimes" />
</ItemGroup>