Nurfog/openccb
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: Mejorar la generación de preguntas con RAG y optimizar el manejo de respuestas de IA

Browse Source
This commit is contained in:
Juan Enrique Allende Cifuentes
2026-04-01 17:13:30 -03:00
parent 1f01a4a36b
commit 7873c0c831
1 changed files with 112 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
services/cms-service/src/handlers_test_templates.rs
+112 -22
View File
@@ -638,10 +638,9 @@ pub async fn generate_questions_with_rag(
Json(payload): Json<RagGenerationPayload>,
) -> Result<Json<Vec<TestTemplateQuestion>>, (StatusCode, String)> {
use common::ai::{self, generate_embedding};
use reqwest::Client;
use serde_json::json;
let mut mysql_questions: Vec<QuestionBankForRAG> = Vec::new();
let mut mysql_questions: Vec<QuestionBankForRAG>;
// If topic is provided, use semantic search; otherwise use course_id filtering
if let Some(topic) = &payload.topic {
@@ -681,14 +680,17 @@ pub async fn generate_questions_with_rag(
1 - (qb.embedding <=> $1::vector) AS similarity
FROM question_bank qb
WHERE qb.organization_id = $2
AND qb.source = 'imported-mysql'
AND ($3::integer IS NULL OR (qb.source_metadata->>'idCursos')::integer = $3)
AND qb.embedding IS NOT NULL
ORDER BY qb.embedding <=> $1::vector
LIMIT $3
LIMIT $4
"#
)
.bind(&pgvector)
.bind(org_ctx.id)
.bind(payload.num_questions.unwrap_or(5) * 3) // Get more for diversity
.bind(payload.course_id)
.bind(payload.num_questions.unwrap_or(5) * 3) // Get more for diversity
.fetch_all(&pool)
.await
.map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, format!("Semantic search failed: {}", e)))?;
@@ -717,16 +719,66 @@ pub async fn generate_questions_with_rag(
) as nivel_curso
FROM question_bank qb
WHERE qb.organization_id = $1
AND qb.question_text ILIKE $2
LIMIT $3
AND qb.source = 'imported-mysql'
AND ($2::integer IS NULL OR (qb.source_metadata->>'idCursos')::integer = $2)
AND (
qb.question_text ILIKE $3
OR COALESCE(qb.options::text, '') ILIKE $3
)
LIMIT $4
"#
)
.bind(org_ctx.id)
.bind(payload.course_id)
.bind(&format!("%{}%", topic))
.bind(payload.num_questions.unwrap_or(5) * 3)
.fetch_all(&pool)
.await
.map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, format!("Keyword search failed: {}", e)))?;
if mysql_questions.is_empty() {
tracing::info!(
"No semantic or keyword matches for topic; falling back to imported MySQL questions for course {:?}",
payload.course_id
);
if let Some(course_id) = payload.course_id {
mysql_questions = sqlx::query_as(
r#"
SELECT
qb.question_text as descripcion,
qb.options,
COALESCE(
(qb.source_metadata->>'idPlanDeEstudios')::integer,
0
) as id_plan_de_estudios,
COALESCE(
qb.source_metadata->>'plan_nombre',
''
) as plan_nombre,
COALESCE(
(qb.source_metadata->>'nivel_curso')::integer,
NULL
) as nivel_curso
FROM question_bank qb
WHERE qb.organization_id = $1
AND qb.source = 'imported-mysql'
AND (qb.source_metadata->>'idCursos')::integer = $2
ORDER BY qb.created_at DESC
"#
)
.bind(org_ctx.id)
.bind(course_id)
.fetch_all(&pool)
.await
.map_err(|e| {
(
StatusCode::INTERNAL_SERVER_ERROR,
format!("Course fallback failed: {}", e),
)
})?;
}
}
}
}
} else if let Some(course_id) = payload.course_id {
@@ -794,8 +846,11 @@ pub async fn generate_questions_with_rag(
.map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, format!("Failed to fetch questions: {}", e)))?;
}
if mysql_questions.is_empty() && payload.course_id.is_some() {
return Err((StatusCode::NOT_FOUND, "No questions found in imported question bank for this course. Please import questions from MySQL first.".to_string()));
if mysql_questions.is_empty() {
return Err((
StatusCode::NOT_FOUND,
"No se encontraron preguntas importadas de MySQL para el curso o tema seleccionado. Importa preguntas del banco MySQL o selecciona un curso con datos disponibles.".to_string(),
));
}
// Determine course_type and level from imported data
@@ -814,12 +869,12 @@ pub async fn generate_questions_with_rag(
// 2. Build RAG context from MySQL questions (lightweight format)
let rag_context: String = mysql_questions
.iter()
.take(15) // Use only first 15 for context
.take(8) // Keep context short so Ollama starts responding sooner behind the HTTPS proxy
.map(|q| format!("* {}", q.descripcion))
.collect::<Vec<_>>()
.join("\n");
tracing::info!("RAG context built with {} questions", mysql_questions.len().min(15));
tracing::info!("RAG context built with {} questions", mysql_questions.len().min(8));
// 3. Call AI to generate new questions based on RAG context (Ollama only)
let base_url = std::env::var("LOCAL_OLLAMA_URL").unwrap_or_else(|_| "http://localhost:11434".to_string());
@@ -838,7 +893,7 @@ pub async fn generate_questions_with_rag(
let topic = payload.topic.clone().unwrap_or_else(|| "English grammar".to_string());
let num_questions = payload.num_questions.unwrap_or(5);
// Simplified system prompt for better performance on slower machines
// Keep the prompt compact so the upstream Ollama proxy can receive the first bytes quickly.
let system_prompt = format!(
r#"You are an English Teacher creating quiz questions.
@@ -860,13 +915,9 @@ pub async fn generate_questions_with_rag(
}}
]
RULES FOR OPTIONS:
- Each option must be ONLY the answer text (1-3 words max)
- Do NOT include letters like "A.", "B.", "a)", "b)"
- Do NOT include "Option 1:", "Answer:", or any prefix
- Just the pure answer text (e.g., "downtown", "Paris", "True")
Skills: reading, listening, speaking, writing. Distribute across all 4."#,
Rules:
- Option text only, no prefixes like A. or 1)
- Skills must be one of: reading, listening, speaking, writing"#,
rag_context,
num_questions,
topic
@@ -888,8 +939,12 @@ pub async fn generate_questions_with_rag(
"content": "Generate the questions in valid JSON format."
}
],
"stream": false,
"format": "json" // Ollama native JSON mode
"stream": true,
"format": "json",
"options": {
"temperature": 0.3,
"num_predict": (num_questions * 160).clamp(160, 900)
}
}));
tracing::info!("Sending request to Ollama (model: {}, prompt length: {} chars)", model, system_prompt.len());
@@ -901,11 +956,46 @@ pub async fn generate_questions_with_rag(
tracing::info!("Ollama response status: {}", response.status());
let response_json: serde_json::Value = response.json().await.map_err(|e| {
tracing::error!("Failed to parse AI response: {}", e);
if !response.status().is_success() {
let status = response.status();
let body = response.text().await.unwrap_or_default();
tracing::error!("Ollama returned non-success status {}: {}", status, body);
return Err((
StatusCode::INTERNAL_SERVER_ERROR,
format!("Ollama returned {}. El proxy de IA agotó el tiempo de espera.", status),
));
}
let response_text = response.text().await.map_err(|e| {
tracing::error!("Failed to read AI stream response: {}", e);
(StatusCode::INTERNAL_SERVER_ERROR, "Invalid AI response".to_string())
})?;
let aggregated_content = response_text
.lines()
.filter_map(|line| {
let trimmed = line.trim();
if trimmed.is_empty() {
return None;
}
serde_json::from_str::<serde_json::Value>(trimmed)
.ok()
.and_then(|chunk| {
chunk.get("message")
.and_then(|m| m.get("content"))
.and_then(|content| content.as_str())
.map(str::to_string)
})
})
.collect::<String>();
let response_json = json!({
"message": {
"content": aggregated_content
}
});
tracing::debug!("Ollama response: {:?}", response_json);
// Parse questions from Ollama response