Improved Prompt
# ============================================================
# STEP 1 — INSTALL REQUIRED LIBRARIES
# ============================================================
# Run these in Google Colab
!pip install -q google-genai
!pip install -q faiss-cpu
%env RETRIEVAL_MODE=faiss
# ============================================================
# STEP 2 — IMPORT LIBRARIES
# ============================================================
import os
import numpy as np
import faiss
import google.genai as genai
from google.colab import userdata
# ============================================================
# STEP 3 — LOAD ENVIRONMENT SETTINGS
# ============================================================
# RETRIEVAL MODES:
#
# "cosine" -> brute-force cosine similarity
# "faiss" -> FAISS vector search
#
# Change this anytime later.
#
# For Render deployment:
# use environment variables.
RETRIEVAL_MODE = os.getenv(
"RETRIEVAL_MODE",
"cosine"
)
print(f"Retrieval mode: {RETRIEVAL_MODE}")
# ============================================================
# STEP 4 — CONFIGURE GEMINI API
# ============================================================
GEMINI_API_KEY = userdata.get("GEMINI_API_KEY")
client = genai.Client(api_key=GEMINI_API_KEY)
# ============================================================
# STEP 5 — CREATE DATASET
# ============================================================
documents = [
# --------------------------------------------------------
# POD FAILURES / DEBUGGING
# --------------------------------------------------------
"CrashLoopBackOff occurs when a container repeatedly crashes after starting.",
"OOMKilled happens when a container exceeds its memory limit.",
"A container may crash due to missing environment variables.",
"Incorrect command or entrypoint can cause container startup failure.",
"Application errors inside the container often lead to restarts.",
"kubectl logs retrieves logs from a running container.",
"kubectl describe pod shows events and state transitions.",
"Liveness probes determine if a container should be restarted.",
"Readiness probes determine if a pod can receive traffic.",
# --------------------------------------------------------
# SCHEDULING
# --------------------------------------------------------
"Pods remain pending if no node satisfies resource requests.",
"Node affinity restricts pods to specific nodes.",
"Taints prevent pods from being scheduled on certain nodes.",
"Tolerations allow pods to be scheduled on tainted nodes.",
# --------------------------------------------------------
# SERVICES
# --------------------------------------------------------
"ClusterIP services expose applications within the cluster.",
"NodePort services expose applications on node IPs.",
"LoadBalancer services expose applications externally.",
"Ingress routes HTTP and HTTPS traffic to services.",
# --------------------------------------------------------
# STORAGE
# --------------------------------------------------------
"PersistentVolumes provide storage independent of pods.",
"PersistentVolumeClaims request storage resources.",
"StorageClasses define dynamic provisioning behavior.",
# --------------------------------------------------------
# DEPLOYMENTS
# --------------------------------------------------------
"Deployments manage replica sets and pod updates.",
"Rolling updates gradually replace old pods with new ones.",
"ReplicaSets maintain a stable number of pod replicas.",
# --------------------------------------------------------
# CONFIGURATION
# --------------------------------------------------------
"ConfigMaps store non-sensitive configuration data.",
"Secrets store sensitive data like passwords and tokens.",
"Environment variables can be injected from ConfigMaps and Secrets.",
# --------------------------------------------------------
# IMAGES / REGISTRY
# --------------------------------------------------------
"ImagePullBackOff occurs when Kubernetes cannot pull the container image.",
"Incorrect image name or tag can cause image pull failures.",
"Private registries require imagePullSecrets for authentication.",
# --------------------------------------------------------
# AUTOSCALING
# --------------------------------------------------------
"Horizontal Pod Autoscaler scales based on CPU or metrics.",
# --------------------------------------------------------
# SECURITY
# --------------------------------------------------------
"RBAC controls access permissions inside Kubernetes.",
"RBAC misconfiguration can block access to resources.",
# --------------------------------------------------------
# NETWORKING
# --------------------------------------------------------
"NetworkPolicies control communication between pods.",
# --------------------------------------------------------
# CLEANUP
# --------------------------------------------------------
"Pods stuck in Terminating state may have finalizers blocking deletion."
]
print(f"Total documents: {len(documents)}")
# ============================================================
# STEP 6 — CREATE SLIDING WINDOW CHUNKS
# ============================================================
# WHY?
# ----
# Preserves neighboring semantic context.
#
# Example:
# sentence1 + sentence2 + sentence3
#
# Then:
# sentence2 + sentence3 + sentence4
WINDOW_SIZE = 3
STRIDE = 1
smart_chunks = []
for i in range(0, len(documents) - WINDOW_SIZE + 1, STRIDE):
chunk = documents[i:i + WINDOW_SIZE]
chunk_text = "\n".join(chunk)
smart_chunks.append(chunk_text)
print(f"Total chunks created: {len(smart_chunks)}")
# ============================================================
# STEP 7 — PREPARE STRUCTURED CHUNK DATA
# ============================================================
prepared_data = []
for i, chunk in enumerate(smart_chunks):
prepared_data.append({
"id": f"chunk_{i}",
"text": chunk
})
print(f"Prepared chunks: {len(prepared_data)}")
# ============================================================
# STEP 8 — CREATE EMBEDDING FUNCTION
# ============================================================
def get_embedding(text):
# response = embed_content(
# model="models/gemini-embedding-001",
# contents=text
# )
# return response["embedding"]
response = client.models.embed_content(
model="models/gemini-embedding-001",
contents=text
)
# The new SDK returns a list of embeddings in 'embeddings'
return response.embeddings[0].values
# ============================================================
# STEP 9 — GENERATE CHUNK EMBEDDINGS
# ============================================================
print("Generating embeddings...")
for item in prepared_data:
embedding = get_embedding(item["text"])
item["embedding"] = embedding
print("Embeddings generated successfully.")
# ============================================================
# STEP 10 — NORMALIZATION FUNCTION
# ============================================================
def normalize(vec):
vec = np.array(vec)
return vec / np.linalg.norm(vec)
# ============================================================
# STEP 11 — COSINE SIMILARITY FUNCTION
# ============================================================
def cosine_similarity(a, b):
a = normalize(a)
b = normalize(b)
return np.dot(a, b)
# ============================================================
# STEP 12 — COSINE RETRIEVAL FUNCTION
# ============================================================
def retrieve_cosine(query, top_k=3, min_score=0.55):
# --------------------------------------------------------
# EMBED QUERY
# --------------------------------------------------------
query_embedding = get_embedding(query)
scores = []
# --------------------------------------------------------
# CALCULATE COSINE SIMILARITY
# --------------------------------------------------------
for item in prepared_data:
similarity = cosine_similarity(
query_embedding,
item["embedding"]
)
scores.append((similarity, item))
# --------------------------------------------------------
# SORT BY SCORE
# --------------------------------------------------------
scores = sorted(
scores,
key=lambda x: x[0],
reverse=True
)
# --------------------------------------------------------
# SIMPLE RE-RANKING
# --------------------------------------------------------
reranked = []
query_words = query.lower().split()
for sim, item in scores:
text = item["text"].lower()
keyword_bonus = sum(
word in text for word in query_words
)
final_score = sim + (0.03 * keyword_bonus)
reranked.append((final_score, item))
# --------------------------------------------------------
# SORT AGAIN AFTER RE-RANKING
# --------------------------------------------------------
reranked = sorted(
reranked,
key=lambda x: x[0],
reverse=True
)
# --------------------------------------------------------
# FILTER LOW SCORES
# --------------------------------------------------------
filtered = [
x for x in reranked
if x[0] >= min_score
]
return filtered[:top_k]
# ============================================================
# STEP 13 — CREATE FAISS EMBEDDING MATRIX
# ============================================================
embedding_matrix = []
for item in prepared_data:
embedding_matrix.append(item["embedding"])
embedding_matrix = np.array(
embedding_matrix,
dtype=np.float32
)
print("Embedding matrix shape:")
print(embedding_matrix.shape)
# ============================================================
# STEP 14 — NORMALIZE EMBEDDINGS FOR FAISS
# ============================================================
# IMPORTANT:
#
# IndexFlatIP uses INNER PRODUCT.
#
# If vectors are normalized:
#
# inner product == cosine similarity
faiss.normalize_L2(embedding_matrix)
# ============================================================
# STEP 15 — CREATE FAISS INDEX
# ============================================================
dimension = embedding_matrix.shape[1]
index = faiss.IndexFlatIP(dimension)
print("FAISS index created.")
# ============================================================
# STEP 16 — ADD EMBEDDINGS TO FAISS INDEX
# ============================================================
index.add(embedding_matrix)
print(f"Total vectors indexed: {index.ntotal}")
# ============================================================
# STEP 17 — FAISS RETRIEVAL FUNCTION
# ============================================================
def retrieve_faiss(query, top_k=3):
# --------------------------------------------------------
# EMBED QUERY
# --------------------------------------------------------
query_embedding = get_embedding(query)
# --------------------------------------------------------
# CONVERT TO NUMPY
# --------------------------------------------------------
query_vector = np.array(
[query_embedding],
dtype=np.float32
)
# --------------------------------------------------------
# NORMALIZE QUERY VECTOR
# --------------------------------------------------------
faiss.normalize_L2(query_vector)
# --------------------------------------------------------
# SEARCH FAISS INDEX
# --------------------------------------------------------
scores, indices = index.search(
query_vector,
top_k
)
# --------------------------------------------------------
# FORMAT RESULTS
# --------------------------------------------------------
results = []
for score, idx in zip(scores[0], indices[0]):
item = prepared_data[idx]
results.append((score, item))
return results
# ============================================================
# STEP 18 — RETRIEVAL ROUTER
# ============================================================
# This decides:
#
# cosine retrieval
# OR
# FAISS retrieval
def retrieve_router(query, top_k=3):
if RETRIEVAL_MODE == "cosine":
return retrieve_cosine(
query=query,
top_k=top_k
)
elif RETRIEVAL_MODE == "faiss":
return retrieve_faiss(
query=query,
top_k=top_k
)
else:
raise ValueError(
f"Invalid retrieval mode: {RETRIEVAL_MODE}"
)
# ============================================================
# STEP 19 — BUILD PROMPT - SOME IMPROVEMENTS
# ============================================================
# WHAT THIS IMPROVES
# -------------------
# ✅ Better grounding
# ✅ Reduced hallucinations
# ✅ Better formatting
# ✅ Better instruction following
# ✅ Cleaner troubleshooting answers
#
# IMPORTANT:
# -----------
# This does NOT improve retrieval itself.
#
# It improves:
# HOW the LLM uses retrieved chunks.
def build_prompt(query, retrieved_chunks):
# --------------------------------------------------------
# BUILD CONTEXT SECTION
# --------------------------------------------------------
context_parts = []
for i, (score, item) in enumerate(retrieved_chunks, start=1):
context_parts.append(
f"""
SOURCE {i}
Relevance Score: {score:.4f}
{item["text"]}
"""
)
context_text = "\n".join(context_parts)
# --------------------------------------------------------
# BUILD FINAL PROMPT
# --------------------------------------------------------
prompt = f"""
You are an expert Kubernetes troubleshooting assistant.
Your job is to answer the user's question ONLY using
the retrieved context provided below.
IMPORTANT RULES:
----------------
1. Use ONLY the retrieved context.
2. Do NOT use outside knowledge.
3. Do NOT invent information.
4. Answer using available context.
If the context is incomplete, explicitly mention that the available
information is limited.
5. If the answer is not present in the context at all,
say:
"I don't know based on the provided context."
6. Keep answers:
- concise
- technically accurate
- well-structured
7. When appropriate:
- use bullet points
- explain causes clearly
- provide troubleshooting guidance
8. If multiple possible causes exist,
list them separately.
9. Prefer information from higher relevance scores.
==================================================
RETRIEVED CONTEXT START
==================================================
{context_text}
==================================================
RETRIEVED CONTEXT END
==================================================
==================================================
USER QUESTION
==================================================
{query}
==================================================
ANSWER
==================================================
"""
return prompt
# ============================================================
# STEP 20 — GENERATE ANSWER USING GEMINI
# ============================================================
def generate_answer(prompt):
# model = genai.GenerativeModel(
# "gemini-3-flash-preview"
# )
# response = model.generate_content(prompt)
response = client.models.generate_content(
#model="models/gemini-3-flash-preview",
#model="models/gemini-2.5-flash",
model="models/gemini-2.5-flash-lite",
contents=prompt)
return response.text
# ============================================================
# STEP 21 — MAIN RAG PIPELINE
# ============================================================
def rag_pipeline(query, top_k=3):
# --------------------------------------------------------
# RETRIEVE CHUNKS
# --------------------------------------------------------
retrieved_chunks = retrieve_router(
query=query,
top_k=top_k
)
# --------------------------------------------------------
# BUILD PROMPT
# --------------------------------------------------------
prompt = build_prompt(
query,
retrieved_chunks
)
# --------------------------------------------------------
# GENERATE ANSWER
# --------------------------------------------------------
answer = generate_answer(prompt)
return answer, retrieved_chunks
# # ============================================================
# # STEP 22 — TEST RETRIEVAL ONLY
# # ============================================================
# test_queries = [
# "Why is my pod crashing?",
# "How to debug Kubernetes logs?",
# "What causes OOMKilled?",
# "How do services work in Kubernetes?",
# "Why is my container restarting repeatedly?"
# ]
# for query in test_queries:
# print("\n" + "=" * 80)
# print(f"QUERY: {query}\n")
# results = retrieve_router(query)
# for score, item in results:
# print(f"Score: {score:.4f}")
# print(item["text"])
# print("-" * 40)
# ============================================================
# STEP 23 — FINAL RAG TEST
# ============================================================
test_queries = [
# --------------------------------------------------------
# DIRECTLY ANSWERABLE
# --------------------------------------------------------
"Why is my pod crashing?",
"How do I debug Kubernetes logs?",
"What causes OOMKilled?",
# --------------------------------------------------------
# MULTI-CAUSE QUESTION
# --------------------------------------------------------
"Why is my container restarting repeatedly?",
# --------------------------------------------------------
# PARTIALLY SUPPORTED
# --------------------------------------------------------
"How does Kubernetes networking work?",
# --------------------------------------------------------
# SHOULD TRIGGER 'I DON'T KNOW'
# --------------------------------------------------------
"How do StatefulSets work?",
"How does etcd replication happen?"
]
for query in test_queries:
print("\n" + "=" * 80)
print(f"QUERY: {query}")
# --------------------------------------------------------
# RUN RAG PIPELINE
# --------------------------------------------------------
answer, sources = rag_pipeline(query)
# --------------------------------------------------------
# PRINT ANSWER
# --------------------------------------------------------
print("\nANSWER:\n")
print(answer)
# --------------------------------------------------------
# PRINT SOURCES
# --------------------------------------------------------
print("\nRETRIEVED SOURCES:\n")
for score, item in sources:
print(f"Score: {score:.4f}")
print(item["text"])
print("-" * 40)
======================================================================================
OUTPUT
======================================================================================
Retrieval mode: faiss
Total documents: 34
Total chunks created: 32
Prepared chunks: 32
Generating embeddings...
Embeddings generated successfully.
Embedding matrix shape:
(32, 3072)
FAISS index created.
Total vectors indexed: 32
================================================================================
QUERY: Why is my pod crashing?
ANSWER:
Your pod might be crashing for several reasons, including:
* **Application errors:** These errors inside the container often lead to restarts.
* **Missing environment variables:** The container may be crashing because it requires environment variables that are not set.
* **Incorrect command or entrypoint:** An incorrect command or entrypoint specified for the container can cause it to fail at startup.
* **OOMKilled:** This occurs when a container exceeds its allocated memory limit.
* **CrashLoopBackOff:** This status indicates that a container is repeatedly crashing after starting.
To troubleshoot, you can use the following `kubectl` commands:
* `kubectl logs`: This command retrieves logs from a running container, which can help identify application errors.
* `kubectl describe pod`: This command shows events and state transitions for the pod, which can provide insights into why it's crashing.
RETRIEVED SOURCES:
Score: 0.7078
CrashLoopBackOff occurs when a container repeatedly crashes after starting.
OOMKilled happens when a container exceeds its memory limit.
A container may crash due to missing environment variables.
----------------------------------------
Score: 0.6856
Application errors inside the container often lead to restarts.
kubectl logs retrieves logs from a running container.
kubectl describe pod shows events and state transitions.
----------------------------------------
Score: 0.6782
A container may crash due to missing environment variables.
Incorrect command or entrypoint can cause container startup failure.
Application errors inside the container often lead to restarts.
----------------------------------------
================================================================================
QUERY: How do I debug Kubernetes logs?
ANSWER:
To debug Kubernetes logs, you can use the `kubectl logs` command to retrieve logs from a running container.
RETRIEVED SOURCES:
Score: 0.7614
Application errors inside the container often lead to restarts.
kubectl logs retrieves logs from a running container.
No comments:
Post a Comment