Spaces:
Sleeping
Sleeping
File size: 12,667 Bytes
eb93f4c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 | """
## Sources
1. Connecting to a Pinecone index: https://docs.pinecone.io/guides/inference/generate-embeddings
2. Creating a text loader from a directory: https://github.com/langchain-ai/langchain/discussions/18559
3. Using an LLM: https://python.langchain.com/docs/tutorials/rag/
4. Evaluation with RAGAS: https://docs.ragas.io/en/stable/getstarted/evals/#analyzing-results
5. Dataset creation: https://www.mongodb.com/developer/products/atlas/evaluate-llm-applications-rag/
## Preparation
First, the packages required for handling connections to the LLM and the vector database must be installed.
Subsequently, the corresponding packages are imported along with the `os` module. This module is utilized to handle the assignments to the environment variables for the API keys."""
from pinecone import Pinecone
import os
from langchain_mistralai import ChatMistralAI
from langchain_openai import ChatOpenAI
from langchain_community.document_loaders import DirectoryLoader
from langchain_community.document_loaders import TextLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_core.messages import HumanMessage, SystemMessage
import pathlib
from langchain.schema import Document
from langchain_community.document_loaders import (
CSVLoader, PyPDFLoader, UnstructuredWordDocumentLoader,
UnstructuredPowerPointLoader, UnstructuredMarkdownLoader,
UnstructuredHTMLLoader, NotebookLoader
)
# -------------------------
# UTF-8 safe Text Loader
# -------------------------
class SafeTextLoader:
"""Loads a text file as a single Document, safely handling UTF-8 decoding errors."""
def __init__(self, file_path):
self.file_path = file_path
def load(self):
try:
with open(self.file_path, "rb") as f: # open in binary mode
raw_bytes = f.read()
text = raw_bytes.decode("utf-8", errors="ignore") # decode safely
return [Document(page_content=text, metadata={"source": str(self.file_path)})]
except Exception as e:
print(f"[Error] Failed to read {self.file_path}: {e}")
return []
# -------------------------
# Loader mapping
# -------------------------
LOADER_MAPPING = {
# Text
".txt": SafeTextLoader,
".json": SafeTextLoader,
".md": UnstructuredMarkdownLoader,
".csv": CSVLoader,
".yaml": SafeTextLoader,
".yml": SafeTextLoader,
# Documents
".pdf": PyPDFLoader,
".docx": UnstructuredWordDocumentLoader,
".pptx": UnstructuredPowerPointLoader,
".html": UnstructuredHTMLLoader,
".htm": UnstructuredHTMLLoader,
# Code / Notebook
".ipynb": NotebookLoader,
".py": SafeTextLoader,
".js": SafeTextLoader,
".sql": SafeTextLoader,
}
import pathlib
CONTEXT_ROOT = pathlib.Path(__file__).parent / "context"
# -------------------------
# Directory loader (recursive)
# -------------------------
def create_text_dir_loader(directory_path: str = ""):
"""
Loads all supported files in the context directory (and subfolders) using the appropriate loader.
- If directory_path is empty -> scans entire 'context' folder recursively.
- If directory_path is given -> scans only that subfolder inside 'context'.
"""
# Resolve the target directory
target_dir = CONTEXT_ROOT / directory_path if directory_path else CONTEXT_ROOT
if not target_dir.exists() or not target_dir.is_dir():
print(f"[Error] Target directory does not exist: {target_dir}")
return []
documents = []
for file_path in target_dir.rglob("*"): # recursive
if not file_path.is_file():
continue
ext = file_path.suffix.lower()
loader_cls = LOADER_MAPPING.get(ext)
if loader_cls is None:
print(f"[Skip] Unsupported file type: {file_path}")
continue
try:
loader = loader_cls(str(file_path))
docs = loader.load()
documents.extend(docs)
print(f"[Loaded] {file_path} ({len(docs)} docs)")
except Exception as e:
print(f"[Error] Failed to load {file_path}: {e}")
print(f"[Done] Finished scanning {target_dir}")
return documents
# -------------------------
# Dataset creation
# -------------------------
def create_dataset(directory_path: str = "context"):
"""
Loads all supported files from the given directory (recursively).
Defaults to 'context' if no path is given.
"""
target_dir = pathlib.Path(directory_path).resolve()
if not target_dir.exists() or not target_dir.is_dir():
print(f"[Error] Target directory does not exist: {target_dir}")
return []
documents = []
for file_path in target_dir.rglob("*"): # recursive
if not file_path.is_file():
continue
ext = file_path.suffix.lower()
loader_cls = LOADER_MAPPING.get(ext)
if loader_cls is None:
print(f"[Skip] Unsupported file type: {file_path}")
continue
try:
loader = loader_cls(str(file_path))
docs = loader.load()
documents.extend(docs)
print(f"[Loaded] {file_path} ({len(docs)} docs)")
except Exception as e:
print(f"[Error] Failed to load {file_path}: {e}")
print(f"[Done] Finished scanning {target_dir}")
print(f"Total documents loaded: {len(documents)}")
return documents
############################################################################################
import time
import os
import math
def prepare_RAG(
pinecone_API,
index_name,
chunk_size=1800,
chunk_overlap=200,
llm_model="gpt-5-nano",
dir_name="context",
info=True
):
import time
from langchain_openai import ChatOpenAI
from langchain_mistralai import ChatMistralAI
from langchain.text_splitter import RecursiveCharacterTextSplitter
from pinecone import Pinecone
if info:
print(f"Prepare RAG with model={llm_model}, dir={dir_name}")
if "gpt" in llm_model:
llm = ChatOpenAI(model=llm_model, streaming=True) # streaming enabled
else:
llm = ChatMistralAI(model=llm_model, streaming=True)
documents = create_dataset(dir_name)
if not documents:
print(f"[Warning] No documents found in directory '{dir_name}'. Using existing Pinecone index without upserting new vectors.")
pc = Pinecone(api_key=pinecone_API)
index = pc.Index(index_name)
return index, pc, llm
text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
all_splits = text_splitter.split_documents(documents)
if not all_splits:
print(f"[Warning] No text chunks were created. Using existing Pinecone index.")
pc = Pinecone(api_key=pinecone_API)
index = pc.Index(index_name)
return index, pc, llm
if info:
print("Number of chunks:", len(all_splits))
pc = Pinecone(api_key=pinecone_API)
index = pc.Index(index_name)
initial_batch_size = 94
min_batch_size = 1
total_chunks = len(all_splits)
records = []
def retry_forever(func, *args, **kwargs):
attempt = 1
while True:
try:
return func(*args, **kwargs)
except Exception as e:
wait_time = min(60, 2 ** min(attempt, 6))
print(f"[Retry] {func.__name__} failed (attempt {attempt}): {e}")
print(f"Sleeping {wait_time} seconds before retry...")
time.sleep(wait_time)
attempt += 1
start_idx = 0
batch_size = initial_batch_size
while start_idx < total_chunks:
end_idx = min(start_idx + batch_size, total_chunks)
batch_splits = all_splits[start_idx:end_idx]
if info:
print(f"Embedding batch {start_idx} → {end_idx - 1} (batch_size={batch_size})")
data = [
{"id": f"vec{idx}", "text": chunk.page_content, "metadata": chunk.metadata or {}}
for idx, chunk in enumerate(batch_splits, start=start_idx)
]
# Attempt embedding
embeddings = retry_forever(
pc.inference.embed,
model="llama-text-embed-v2",
inputs=[d['text'] for d in data],
parameters={"input_type": "passage", "truncate": "END"}
)
# Prepare records for upsert
batch_records = [
{"id": d['id'], "values": e['values'], "metadata": {"text": d['text'], **d.get('metadata', {})}}
for d, e in zip(data, embeddings)
]
# Dynamic upsert with batch size adjustment
while batch_records:
try:
retry_forever(index.upsert, vectors=batch_records, namespace="example-namespace")
records.extend(batch_records)
break # success → exit loop
except Exception as e:
if "Request size" in str(e) or "exceeds the maximum" in str(e):
if len(batch_records) == 1:
print(f"[Error] Single vector too large to upsert: {batch_records[0]['id']}")
break
# Reduce batch size by half
batch_records = batch_records[:len(batch_records) // 2]
print(f"[Warning] Batch too large. Reducing batch size to {len(batch_records)} and retrying upsert...")
time.sleep(1) # small delay before retry
else:
raise e
start_idx += batch_size # move to next batch
# Optionally increase batch size gradually if previous upsert succeeded
batch_size = min(initial_batch_size, batch_size * 2)
if info:
print(f"Completed upsert of {len(records)} vectors.")
return index, pc, llm
def retrieve_RAG(prompt_message, pc, index, top_k=5, info=True):
if info:
print("Retrieve RAG with", prompt_message, pc, index, top_k)
"""## Retrieval
The user query is embedded using the same model (specified in the `model` parameter of the `pc.inference.embed()` function) as for embedding the chunks originally upserted into the vector database.
"""
query_embedding = pc.inference.embed(
model="llama-text-embed-v2",
inputs=[prompt_message],
parameters={
"input_type": "query"
}
)
"""The relevant chunks are retrieved using semantic search with the cosine distance similarity measure. The number of chunks to be retrieved is passed in the `top_k` variable."""
retrieved_chunks_raw = index.query(
namespace="example-namespace",
vector=query_embedding[0].values,
top_k=top_k,
include_values=False,
include_metadata=True
)
"""The result of the retrieval is processed, so that the variable `retrieved_chunks` contains a list of the corresponding text splits.
"""
retrieved_chunks = []
for match in retrieved_chunks_raw.matches:
retrieved_chunks.append({
"text": match.metadata.get("text", ""),
"source": match.metadata.get("source", "")
})
return retrieved_chunks
def generate_RAG(prompt_message, llm, retrieved_chunks, info=True):
if info:
print("Generate RAG with", prompt_message, llm)
"""## Generation
The prompt is sent to the LLM together with the context consisting of the data considered most relevant according to the semantic search. These data are stored in the variable `retrieved_chunks`. The part of the prompt containing the explaination of the task to be performed on the provided context is in the variable `prompt_message`.
"""
context_message = "You are an expert in the field of the documents that are loaded. Use only the these documents and source files as the context. If you don't know the answer or the answer is not included in the context, then do not answer."
prompt = [HumanMessage(prompt_message),
SystemMessage(context_message),
HumanMessage(str(retrieved_chunks))]
"""The prompt is sent to the selected LLM using the `invoke()` function. As a return value, this function passes the respective answer of the LLM."""
response = llm.invoke(prompt)
return response
|