Un chatbot para reparación de flotas como ejemplo de RAG en lenguaje sencillo con Chroma y Ollama
RAG explicado mediante un chatbot para reparación de flotas: los informes de reparación de automóviles llegan por WhatsApp, y Chroma, los embeddings de MiniLM y Llama 3.2 en Ollama realizan la primera evaluación.
Fuentes: Understanding RAG in simplified manner with an actual use case
RAG es más fácil de comprender a través de un caso concreto que mediante diagramas, y una flota de automóviles pequeña constituye un buen ejemplo. Imagine a un operador cuyos vehículos son casi todos del mismo modelo, un Daihatsu Mira, y que cuenta con un manual detallado que abarca numerosos defectos de reparación. Cuando un conductor informa de un problema, ese manual suele permitir al operador solucionar rápidamente los síntomas más comunes. Aun así, el mecánico debe inspeccionar el automóvil, pero desde la perspectiva de una IA agente resulta atractivo dejar que el sistema de recuperación de información se encargue de la primera fase de diagnóstico.
El escenario
Los conductores informan de sus problemas a un chatbot a través de WhatsApp. El mensaje llega a una API mediante un webhook y es procesado; en algún punto intermedio, un agente automatizado genera un diagnóstico con una posible sugerencia, la cual se registra como resultado del diagnóstico.
Un ejemplo de intercambio: el conductor escribe que los frenos emiten un sonido estridente al presionarlos y que necesitan ser revisados. Un diagnóstico automático podría verse así:
The screeching noise when pressing the brake pedal is a symptom that requires attention.
Possible causes
- The screeching noise may be caused by:
• Glazed or contaminated brake-pad friction material
• Loose, damaged, incorrectly installed, or missing pad hardware and shims
• Dry or seized caliper slide pins
• Debris, corrosion, scoring, or an uneven surface on the brake rotor
• Normal vibration characteristics of some pad and rotor combinations
Diagnostic checks
• Confirm when the sound occurs during a cautious test in a safe area
• Identify the wheel area producing the noise
• Inspect both the inner and outer brake pads for glazing, contamination, cracking, or uneven wear
• Check whether a wear indicator is touching the rotor and inspect all clips, shims, and mounting hardware
• Inspect the rotor for embedded debris, scoring, heat spots, corrosion, and excessive runout or thickness variation
• Check that the caliper piston and slide pins move correctly and inspect for damaged boots or brake-fluid leakage
Safety
Noise alone does not prove one cause. If the noise becomes grinding, braking performance is reduced, the vehicle pulls while braking, a brake warning light is illuminated, or brake fluid is leaking, stop driving and obtain qualified help immediately.
Dónde encaja RAG
Los manuales de reparación sirven como conocimiento diagnóstico. Uno de ellos es un PDF sobre pastillas de freno. Cuando llega un mensaje del conductor, asume el control la pipeline de RAG; el modelo utilizado aquí es Llama 3.2, proporcionado a través de Ollama, Chroma DB para el almacenamiento vectorial, y el modelo all-MiniLM-L6-v2, a través de SentenceTransformerEmbeddingFunction, para las representaciones vectoriales.
El flujo se desarrolla en siete pasos:
- El conductor envía un mensaje a través del chatbot.
- El mensaje llega al webhook.
- Se activa la pipeline de RAG.
- Primero se realiza la recuperación de información: se leen los manuales en PDF y se dividen en fragmentos, se adjuntan metadatos como el número de página y el nombre del manual, cada fragmento se procesa con la función sentence-transformer y los vectores se almacenan en Chroma DB.
- El mensaje recibido se utiliza entonces para consultar a Chroma y obtener el contexto relevante para este tipo de pregunta de reparación.
- La instrucción se enriquece con el contexto recuperado junto con la pregunta.
- La instrucción enriquecida se envía al LLM, y su respuesta vuelve al usuario.
Un ejemplo de script en Python para el manual de mantenimiento de pastillas de freno implementa todo esto:
"""A small, real RAG pipeline for the brake-pad maintenance manual."""
import os
import re
from pathlib import Path
os.environ["TOKENIZERS_PARALLELISM"] = "false"
import chromadb
from chromadb.utils.embedding_functions import SentenceTransformerEmbeddingFunction
from langchain_ollama import OllamaLLM
from langchain_text_splitters import RecursiveCharacterTextSplitter
from pypdf import PdfReader
PDF_PATH = Path(__file__).with_name("brake_pad_maintenance_manual.pdf")
USER_QUERY = "My brakes are making screeching noises when I press the pedal."
def ingest_pdf(collection, pdf_path: Path) -> None:
"""Extract, chunk, and add a real PDF to ChromaDB."""
if not pdf_path.exists():
raise FileNotFoundError(f"Manual not found: {pdf_path}")
splitter = RecursiveCharacterTextSplitter(chunk_size=1200, chunk_overlap=150)
chunks = []
metadatas = []
ids = []
for page_number, page in enumerate(PdfReader(pdf_path).pages, start=1):
page_text = re.sub(r"\s+", " ", page.extract_text() or "").strip()
for chunk_number, chunk in enumerate(splitter.split_text(page_text)):
if not chunk.strip():
continue
chunks.append(chunk)
metadatas.append(
{
"source_file": pdf_path.name,
"page": page_number,
}
)
ids.append(f"page_{page_number}_chunk_{chunk_number}")
if not chunks:
raise ValueError(f"No text could be extracted from {pdf_path.name}.")
collection.add(documents=chunks, metadatas=metadatas, ids=ids)
print(f"Ingested {len(chunks)} chunks from {pdf_path.name}.\n")
def generate_llm_response(prompt: str) -> str:
"""Generate a grounded answer with the local Ollama model."""
return OllamaLLM(model="llama3.2", temperature=0).invoke(prompt)
def run_rag_pipeline(collection, user_query: str):
"""Retrieve relevant PDF passages, augment the prompt, and generate."""
results = collection.query(
query_texts=[user_query],
n_results=min(2, collection.count()),
)
retrieved_chunks = results["documents"][0]
retrieved_metadata = results["metadatas"][0]
if not retrieved_chunks:
raise ValueError("No relevant manual passages were retrieved.")
context = "\n\n".join(
f"[{source['source_file']}, page {source['page']}]\n{chunk}"
for chunk, source in zip(retrieved_chunks, retrieved_metadata)
)
augmented_prompt = f"""
You are a cautious fleet mechanic assistant.
Answer strictly from the supplied manual passages in no more than 180 words.
For a reported symptom, use these four headings: Assessment, Possible causes,
Diagnostic checks, and Safety. Make clear that noise alone does not prove one
cause. Cover the relevant checks in concise bullets without claiming that an
inspection has already been performed or asking the user to continue. Do not
mention a measurement unless it appears in the context and directly helps
answer the question. Cite the source file and page number. If the answer is
absent, say it was not found in the manual.
=== MANUAL CONTEXT ===
{context}
=== USER QUERY ===
{user_query}
=== ANSWER ===
"""
response = generate_llm_response(augmented_prompt).rstrip()
citations = sorted(
{
f"{source['source_file']}, page {source['page']}"
for source in retrieved_metadata
}
)
response = f"{response}\n\nSources: {'; '.join(citations)}"
return response, augmented_prompt
def main() -> None:
embedding_function = SentenceTransformerEmbeddingFunction(
model_name="all-MiniLM-L6-v2"
)
collection = chromadb.Client().create_collection(
name="brake_pad_manual_rag",
embedding_function=embedding_function,
)
ingest_pdf(collection, PDF_PATH)
response, assembled_prompt = run_rag_pipeline(collection, USER_QUERY)
print("=== AUGMENTED PROMPT SENT TO LLM ===")
print(assembled_prompt)
print("=== FINAL GENERATED RESPONSE ===")
print(response)
if __name__ == "__main__":
main()
El manual de demostración que se muestra es un breve documento de capacitación:
Demo Brake-Pad Maintenance Manual
Training document only. Brake systems are safety-critical. Always use the vehicle manufacturer's service
manual, specified service limits, approved lifting points, and torque values.
Brake-pad inspection
1. Park on a level surface, prevent vehicle movement, and lift the vehicle only at approved points. Support it
securely with rated stands.
2. Remove the wheel using the vehicle manufacturer's procedure.
3. Inspect the inner and outer pads. Check friction material at several points and compare it with the vehiclespecific
minimum thickness.
4. Replace pads that are at or below their service limit, contaminated, cracked, damaged, or wearing unevenly.
5. Inspect the rotor for scoring, heat damage, cracks, and excessive corrosion. Measure rotor thickness and
runout against vehicle-specific limits.
6. Check the caliper, slide pins, dust boots, and brake hose for seizure, damage, or fluid leakage.
Symptom: screeching when the brake pedal is pressed
A high-pitched screech or squeal during braking can indicate that a brake-pad wear indicator is contacting the
rotor, but noise alone does not prove that the pads are worn out.
Other possible causes
Glazed or contaminated brake-pad friction material.
Loose, damaged, incorrectly installed, or missing pad hardware and shims.
Dry or seized caliper slide pins that cause uneven pad contact.
Debris, corrosion, scoring, or an uneven surface on the brake rotor.
Normal vibration characteristics of some pad and rotor combinations.
Possible diagnostic procedure
1. Confirm when the sound occurs during a cautious test in a safe area. Note whether it occurs only while
braking, at low or high speed, or continuously while the wheel rotates.
2. Identify the wheel area producing the noise. Do not assume that the loudest apparent location is the exact
source.
3. Inspect both the inner and outer brake pads. Compare their thickness with the vehicle-specific service limit
and look for glazing, contamination, cracking, or uneven wear.
4. Check whether a wear indicator is touching the rotor and inspect all clips, shims, and mounting hardware.
5. Inspect the rotor for embedded debris, scoring, heat spots, corrosion, and excessive runout or thickness
variation.
6. Check that the caliper piston and slide pins move correctly and inspect for damaged boots or brake-fluid
leakage.
7. Correct the confirmed cause using the vehicle manufacturer's procedure, then verify the repair with a
cautious low-speed brake test.
Stop driving and obtain qualified help if the noise becomes grinding, braking performance is reduced, the
vehicle pulls while braking, a brake warning light is illuminated, or brake fluid is leaking.
Brake-pad replacement
1. Follow the model-specific caliper removal procedure. Never let the caliper hang from the brake hose.
2. Retract the caliper piston using the correct tool and procedure. Put an electronic parking brake into service
mode when the manufacturer requires it.
3. Clean the pad supports and inspect the hardware. Apply only manufacturer-approved brake lubricant to
specified contact points. Keep lubricant away from pad and rotor friction surfaces.
4. Install new pads as a complete axle set. Replace clips, shims, or wear sensors when required.
5. Reinstall the caliper and wheel. Tighten every fastener to the vehicle manufacturer's torque specification.
6. Before moving the vehicle, pump the brake pedal until it is firm. Check the brake-fluid level and inspect for
leaks.
7. Follow the pad manufacturer's bedding procedure, then perform a cautious low-speed function test in a safe
area.
Stop and obtain qualified help if
The pedal remains soft, brake fluid is leaking, the caliper is damaged or seized, the rotor is below its service limit,
or any required specification is unavailable.
Y las dependencias se encuentran en un archivo de requisitos:
scikit-learn
pydantic
langchain
langchain-community
langchain-huggingface
langchain-ollama
langchain-text-splitters
chromadb
pypdf
sentence-transformers
Lección relacionada
- Por qué RAG falla con documentos reales: particionamiento, búsqueda híbrida, reclasificación y evaluación — Patrones básicos para un pipeline RAG en producción: diseño en dos fases, opciones de particionamiento, embeddings y almacenes vectoriales, búsqueda híbrida, reclasificación y evaluación de la recuperación.
- Auditoría de la recuperación RAG para chunks obsoletos, desviaciones y contexto ruidoso — Un script de auditoría en Python para la recuperación RAG: consultas de referencia, control de relevancia mediante coseno, verificaciones de desviación Kolmogorov-Smirnov y eficiencia del contexto, además de cómo leer su informe.