Spaces:
Running
Running
File size: 7,831 Bytes
11c72a2 | 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 | import hashlib
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
from sentence_transformers import SentenceTransformer
import faiss
from langchain.prompts import ChatPromptTemplate
from langchain.docstore.document import Document
from langchain.memory import ConversationBufferMemory
from langchain.chains import ConversationChain
import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"
# --- MMR Utilities ---
def build_mmr_index(docs):
texts = [doc['text'] for doc in docs if 'text' in doc]
documents = [Document(page_content=text) for text in texts]
model = SentenceTransformer("all-MiniLM-L6-v2")
embeddings = model.encode([doc.page_content for doc in documents], convert_to_numpy=True)
faiss.normalize_L2(embeddings)
index = faiss.IndexFlatIP(embeddings.shape[1])
index.add(embeddings)
return model, index, embeddings, documents
def get_mmr_sample(model, index, embeddings, documents, query, k=15, lambda_mult=0.7):
if len(documents) == 0:
print("Warning: No documents available, returning empty list.")
return []
if len(documents) <= k:
print(f"Warning: Only {len(documents)} documents available, returning all.")
return documents
else:
query_vec = model.encode(query, convert_to_numpy=True)
query_vec = query_vec / np.linalg.norm(query_vec)
# Get candidate indices from FAISS (k * 4 or less if not enough documents)
num_candidates = min(k * 4, len(documents))
D, I = index.search(np.expand_dims(query_vec, axis=0), num_candidates)
candidate_idxs = list(I[0])
selected = []
while len(selected) < k and candidate_idxs:
if not selected:
selected.append(candidate_idxs.pop(0))
continue
mmr_scores = []
for idx in candidate_idxs:
relevance = cosine_similarity([query_vec], [embeddings[idx]])[0][0]
diversity = max([
cosine_similarity([embeddings[idx]], [embeddings[sel]])[0][0]
for sel in selected
])
mmr_score = lambda_mult * relevance - (1 - lambda_mult) * diversity
mmr_scores.append((idx, mmr_score))
next_best = max(mmr_scores, key=lambda x: x[1])[0]
selected.append(next_best)
candidate_idxs.remove(next_best)
return [documents[i] for i in selected]
# --- Summarization ---
def summarize_docs(word, timestamp, docs, llm, k):
if not docs:
return "No documents available for this word at this time.", [], 0
try:
model, index, embeddings, documents = build_mmr_index(docs)
mmr_docs = get_mmr_sample(model, index, embeddings, documents, query=word, k=k)
context_texts = "\n".join(f"- {doc.page_content}" for doc in mmr_docs)
prompt_template = ChatPromptTemplate.from_template(
"Given the following documents from {timestamp} containing the word '{word}', "
"identify the key themes or distinct discussion points that were prevalent during that time. "
"Do NOT describe each bullet in detail. Be concise. Each bullet should be a short phrase or sentence "
"capturing a unique, non-overlapping theme. Avoid any elaboration, examples, or justification.\n\n"
"Return no more than 5–7 bullets.\n\n"
"{context_texts}\n\nSummary:"
)
chain = prompt_template | llm
summary = chain.invoke({
"word": word,
"timestamp": timestamp,
"context_texts": context_texts
}).content.strip()
return summary, mmr_docs
except Exception as e:
return f"[Error summarizing: {e}]", [], 0
def summarize_multiword_docs(words, timestamp, docs, llm, k):
if not docs:
return "No common documents available for these words at this time.", []
try:
model, index, embeddings, documents = build_mmr_index(docs)
query = " ".join(words)
mmr_docs = get_mmr_sample(model, index, embeddings, documents, query=query, k=k)
context_texts = "\n".join(f"- {doc.page_content}" for doc in mmr_docs)
prompt_template = ChatPromptTemplate.from_template(
"Given the following documents from {timestamp} that all mention the words: '{word_list}', "
"identify the key themes or distinct discussion points that were prevalent during that time. "
"Do NOT describe each bullet in detail. Be concise. Each bullet should be a short phrase or sentence "
"capturing a unique, non-overlapping theme. Avoid any elaboration, examples, or justification.\n\n"
"Return no more than 5–7 bullets.\n\n"
"{context_texts}\n\n"
"Concise Thematic Summary:"
)
chain = prompt_template | llm
summary = chain.invoke({
"word_list": ", ".join(words),
"timestamp": timestamp,
"context_texts": context_texts
}).content.strip()
return summary, mmr_docs
except Exception as e:
return f"[Error summarizing: {e}]", []
# --- Follow-up Question Handler (Improved) ---
def ask_multiturn_followup(history: list, question: str, llm, context_texts: str) -> str:
"""
Handles multi-turn follow-up questions based on a provided set of documents.
This function now REQUIRES context_texts to be provided, ensuring the LLM
is always grounded in the source documents for follow-up questions.
Args:
history (list): A list of dictionaries representing the conversation history
(e.g., [{"role": "user", "content": "..."}]).
question (str): The user's new follow-up question.
llm: The initialized language model instance.
context_texts (str): A single string containing all the numbered documents
for context.
Returns:
str: The AI's response to the follow-up question.
"""
try:
# 1. Reconstruct conversation memory from the history provided from the UI
memory = ConversationBufferMemory(return_messages=True)
for turn in history:
if turn["role"] == "user":
memory.chat_memory.add_user_message(turn["content"])
elif turn["role"] == "assistant":
memory.chat_memory.add_ai_message(turn["content"])
# 2. Define the system instruction that grounds the LLM
system_instruction = (
"You are an assistant answering questions strictly based on the provided sample documents below. "
"Your memory contains the previous turns of this conversation. "
"If the answer is not clearly available in the text, respond with: "
"'The information is not available in the documents provided.'\n\n"
)
# 3. Create the full prompt. No more conditional logic, as context is required.
# The `ConversationChain` will automatically use the memory, so we only need
# to provide the current input, which includes the grounding documents.
full_prompt = (
f"{system_instruction}"
f"--- DOCUMENTS ---\n{context_texts.strip()}\n\n"
f"--- QUESTION ---\n{question}"
)
# 4. Create and run the conversation chain
conversation = ConversationChain(llm=llm, memory=memory, verbose=False)
response = conversation.predict(input=full_prompt)
return response.strip()
except Exception as e:
# Good practice to log the full exception for easier debugging
print(f"[ERROR] in ask_multiturn_followup: {e}")
return f"[Error during multi-turn follow-up. Please check the logs.]" |