Arthur Passuello
Test to fix prompt
118e3aa
"""
HuggingFace API-based answer generation for deployment environments.
This module provides answer generation using HuggingFace's Inference API,
optimized for cloud deployment where local LLMs aren't feasible.
"""
import json
import logging
from dataclasses import dataclass
from typing import List, Dict, Any, Optional, Generator, Tuple
from datetime import datetime
import re
from pathlib import Path
import requests
import os
import sys
# Import technical prompt templates
from .prompt_templates import TechnicalPromptTemplates
logger = logging.getLogger(__name__)
@dataclass
class Citation:
"""Represents a citation to a source document chunk."""
chunk_id: str
page_number: int
source_file: str
relevance_score: float
text_snippet: str
@dataclass
class GeneratedAnswer:
"""Represents a generated answer with citations."""
answer: str
citations: List[Citation]
confidence_score: float
generation_time: float
model_used: str
context_used: List[Dict[str, Any]]
class HuggingFaceAnswerGenerator:
"""
Generates answers using HuggingFace Inference API with hybrid reliability.
🎯 HYBRID APPROACH - Best of Both Worlds:
- Primary: High-quality open models (Zephyr-7B, Mistral-7B-Instruct)
- Fallback: Reliable classics (DialoGPT-medium)
- Foundation: HF GPT's proven Docker + auth setup
- Pro Benefits: Better rate limits, priority processing
Optimized for deployment environments with:
- Fast API-based inference
- No local model requirements
- Citation extraction and formatting
- Confidence scoring
- Automatic fallback for reliability
"""
def __init__(
self,
model_name: str = "sshleifer/distilbart-cnn-12-6",
api_token: Optional[str] = None,
temperature: float = 0.3,
max_tokens: int = 512
):
"""
Initialize the HuggingFace answer generator.
Args:
model_name: HuggingFace model to use
api_token: HF API token (optional, uses free tier if None)
temperature: Generation temperature (0.0-1.0)
max_tokens: Maximum tokens to generate
"""
self.model_name = model_name
# Try multiple common token environment variable names
self.api_token = (api_token or
os.getenv("HUGGINGFACE_API_TOKEN") or
os.getenv("HF_TOKEN") or
os.getenv("HF_API_TOKEN"))
self.temperature = temperature
self.max_tokens = max_tokens
# Hybrid approach: Classic API + fallback models
self.api_url = f"https://api-inference.huggingface.co/models/{model_name}"
# Prepare headers
self.headers = {"Content-Type": "application/json"}
if self.api_token:
self.headers["Authorization"] = f"Bearer {self.api_token}"
logger.info("Using authenticated HuggingFace API")
else:
logger.info("Using free HuggingFace API (rate limited)")
# Only include models that actually work based on tests
self.fallback_models = [
"deepset/roberta-base-squad2", # Q&A model - perfect for RAG
"sshleifer/distilbart-cnn-12-6", # Summarization - also good
"facebook/bart-base", # Base BART - works but needs right format
]
def _call_api_with_model(self, prompt: str, model_name: str) -> str:
"""Call API with a specific model (for fallback support)."""
fallback_url = f"https://api-inference.huggingface.co/models/{model_name}"
# SIMPLIFIED payload that works
payload = {"inputs": prompt}
response = requests.post(
fallback_url,
headers=self.headers,
json=payload,
timeout=30
)
response.raise_for_status()
result = response.json()
# Handle response
if isinstance(result, list) and len(result) > 0:
if isinstance(result[0], dict):
return result[0].get("generated_text", "").strip()
else:
return str(result[0]).strip()
elif isinstance(result, dict):
return result.get("generated_text", "").strip()
else:
return str(result).strip()
def _create_system_prompt(self) -> str:
"""Create system prompt optimized for the model type."""
if "squad" in self.model_name.lower() or "roberta" in self.model_name.lower():
# RoBERTa Squad2 uses question/context format - no system prompt needed
return ""
elif "gpt2" in self.model_name.lower() or "distilgpt2" in self.model_name.lower():
# GPT-2 style completion prompt - simpler is better
return "Based on the following context, answer the question.\n\nContext: "
elif "llama" in self.model_name.lower():
# Llama-2 chat format
return """<s>[INST] You are a helpful technical documentation assistant. Answer the user's question based only on the provided context. Always cite sources using [chunk_X] format.
Context:"""
elif "flan" in self.model_name.lower() or "t5" in self.model_name.lower():
# Flan-T5 instruction format - simple and direct
return """Answer the question based on the context below. Cite sources using [chunk_X] format.
Context: """
elif "falcon" in self.model_name.lower():
# Falcon instruction format
return """### Instruction: Answer based on the context and cite sources with [chunk_X].
### Context: """
elif "bart" in self.model_name.lower():
# BART summarization format
return """Summarize the answer to the question from the context. Use [chunk_X] for citations.
Context: """
else:
# Default instruction prompt for other models
return """You are a technical documentation assistant that provides clear, accurate answers based on the provided context.
CORE PRINCIPLES:
1. ANSWER DIRECTLY: If context contains the answer, provide it clearly and confidently
2. BE CONCISE: Keep responses focused and avoid unnecessary uncertainty language
3. CITE ACCURATELY: Use [chunk_X] citations for every fact from context
RESPONSE GUIDELINES:
- If context has sufficient information → Answer directly and confidently
- If context has partial information → Answer what's available, note what's missing briefly
- If context is irrelevant → Brief refusal: "This information isn't available in the provided documents"
CITATION FORMAT:
- Use [chunk_1], [chunk_2] etc. for all facts from context
- Example: "According to [chunk_1], RISC-V is an open-source architecture."
Be direct, confident, and accurate. If the context answers the question, provide that answer clearly."""
def _format_context(self, chunks: List[Dict[str, Any]]) -> str:
"""
Format retrieved chunks into context for the LLM.
Args:
chunks: List of retrieved chunks with metadata
Returns:
Formatted context string
"""
context_parts = []
for i, chunk in enumerate(chunks):
chunk_text = chunk.get('content', chunk.get('text', ''))
page_num = chunk.get('metadata', {}).get('page_number', 'unknown')
source = chunk.get('metadata', {}).get('source', 'unknown')
context_parts.append(
f"[chunk_{i+1}] (Page {page_num} from {source}):\n{chunk_text}\n"
)
return "\n---\n".join(context_parts)
def _call_api(self, prompt: str) -> str:
"""
Call HuggingFace Inference API.
Args:
prompt: Input prompt for the model
Returns:
Generated text response
"""
# Validate prompt
if not prompt or len(prompt.strip()) < 5:
logger.warning(f"Prompt too short: '{prompt}' - padding it")
prompt = f"Please provide information about: {prompt}. Based on the context, give a detailed answer."
# Model-specific payload formatting
if "squad" in self.model_name.lower() or "roberta" in self.model_name.lower():
# RoBERTa Squad2 needs question and context separately
# Parse the structured prompt format we create
if "Context:" in prompt and "Question:" in prompt:
# Split by the markers we use
parts = prompt.split("Question:")
if len(parts) == 2:
context_part = parts[0].replace("Context:", "").strip()
question_part = parts[1].strip()
else:
# Fallback
question_part = "What is this about?"
context_part = prompt
else:
# Fallback for unexpected format
question_part = "What is this about?"
context_part = prompt
# Clean up the context and question
context_part = context_part.replace("---", "").strip()
if not question_part or len(question_part.strip()) < 3:
question_part = "What is the main information?"
# Debug output
print(f"🔍 Squad2 Question: {question_part[:100]}...")
print(f"🔍 Squad2 Context: {context_part[:200]}...")
payload = {
"inputs": {
"question": question_part,
"context": context_part
}
}
elif "bart" in self.model_name.lower() or "distilbart" in self.model_name.lower():
# BART/DistilBART for summarization
if len(prompt) < 50:
prompt = f"{prompt} Please provide a comprehensive answer based on the available information."
payload = {
"inputs": prompt,
"parameters": {
"max_length": 150,
"min_length": 10,
"do_sample": False
}
}
else:
# Simple payload for other models
payload = {"inputs": prompt}
try:
logger.info(f"Calling API URL: {self.api_url}")
logger.info(f"Headers: {self.headers}")
logger.info(f"Payload: {payload}")
response = requests.post(
self.api_url,
headers=self.headers,
json=payload,
timeout=30
)
logger.info(f"Response status: {response.status_code}")
logger.info(f"Response headers: {response.headers}")
if response.status_code == 503:
# Model is loading, wait and retry
logger.warning("Model loading, waiting 20 seconds...")
import time
time.sleep(20)
response = requests.post(
self.api_url,
headers=self.headers,
json=payload,
timeout=30
)
logger.info(f"Retry response status: {response.status_code}")
if response.status_code == 404:
logger.error(f"Model not found: {self.model_name}")
logger.error(f"Response text: {response.text}")
# Try fallback models
for fallback_model in self.fallback_models:
if fallback_model != self.model_name:
logger.info(f"Trying fallback model: {fallback_model}")
try:
return self._call_api_with_model(prompt, fallback_model)
except Exception as e:
logger.warning(f"Fallback model {fallback_model} failed: {e}")
continue
return "All models are currently unavailable. Please try again later."
response.raise_for_status()
result = response.json()
# Handle different response formats based on model type
print(f"🔍 API Response type: {type(result)}")
print(f"🔍 API Response preview: {str(result)[:300]}...")
if isinstance(result, dict) and "answer" in result:
# RoBERTa Squad2 format: {"answer": "...", "score": ..., "start": ..., "end": ...}
answer = result["answer"].strip()
print(f"🔍 Squad2 extracted answer: {answer}")
return answer
elif isinstance(result, list) and len(result) > 0:
# Check for DistilBART format (returns dict with summary_text)
if isinstance(result[0], dict) and "summary_text" in result[0]:
return result[0]["summary_text"].strip()
# Check for nested list (BART format: [[...]])
elif isinstance(result[0], list) and len(result[0]) > 0:
if isinstance(result[0][0], dict):
return result[0][0].get("summary_text", str(result[0][0])).strip()
else:
# BART base returns embeddings - not useful for text generation
logger.warning("BART returned embeddings instead of text")
return "Model returned embeddings instead of text. Please try a different model."
# Regular list format
elif isinstance(result[0], dict):
# Try different keys that models might use
text = (result[0].get("generated_text", "") or
result[0].get("summary_text", "") or
result[0].get("translation_text", "") or
result[0].get("answer", "") or
str(result[0]))
# Remove the input prompt from the output if present
if isinstance(prompt, str) and text.startswith(prompt):
text = text[len(prompt):].strip()
return text
else:
return str(result[0]).strip()
elif isinstance(result, dict):
# Some models return dict directly
text = (result.get("generated_text", "") or
result.get("summary_text", "") or
result.get("translation_text", "") or
result.get("answer", "") or
str(result))
# Remove input prompt if model included it
if isinstance(prompt, str) and text.startswith(prompt):
text = text[len(prompt):].strip()
return text
elif isinstance(result, str):
return result.strip()
else:
logger.error(f"Unexpected response format: {type(result)} - {result}")
return "I apologize, but I couldn't generate a response."
except requests.exceptions.RequestException as e:
logger.error(f"API request failed: {e}")
if hasattr(e, 'response') and e.response is not None:
logger.error(f"Response status: {e.response.status_code}")
logger.error(f"Response body: {e.response.text}")
return f"API Error: {str(e)}. Using free tier? Try adding an API token."
except Exception as e:
logger.error(f"Unexpected error: {e}")
import traceback
logger.error(f"Traceback: {traceback.format_exc()}")
return f"Error: {str(e)}. Please check logs for details."
def _extract_citations(self, answer: str, chunks: List[Dict[str, Any]]) -> Tuple[str, List[Citation]]:
"""
Extract citations from the generated answer and integrate them naturally.
Args:
answer: Generated answer with [chunk_X] citations
chunks: Original chunks used for context
Returns:
Tuple of (natural_answer, citations)
"""
citations = []
citation_pattern = r'\[chunk_(\d+)\]'
cited_chunks = set()
# Find [chunk_X] citations and collect cited chunks
matches = re.finditer(citation_pattern, answer)
for match in matches:
chunk_idx = int(match.group(1)) - 1 # Convert to 0-based index
if 0 <= chunk_idx < len(chunks):
cited_chunks.add(chunk_idx)
# FALLBACK: If no explicit citations found but we have an answer and chunks,
# create citations for the top chunks that were likely used
if not cited_chunks and chunks and len(answer.strip()) > 50:
# Use the top chunks that were provided as likely sources
num_fallback_citations = min(3, len(chunks)) # Use top 3 chunks max
cited_chunks = set(range(num_fallback_citations))
print(f"🔧 HF Fallback: Creating {num_fallback_citations} citations for answer without explicit [chunk_X] references", file=sys.stderr, flush=True)
# Create Citation objects for each cited chunk
chunk_to_source = {}
for idx in cited_chunks:
chunk = chunks[idx]
citation = Citation(
chunk_id=chunk.get('id', f'chunk_{idx}'),
page_number=chunk.get('metadata', {}).get('page_number', 0),
source_file=chunk.get('metadata', {}).get('source', 'unknown'),
relevance_score=chunk.get('score', 0.0),
text_snippet=chunk.get('content', chunk.get('text', ''))[:200] + '...'
)
citations.append(citation)
# Map chunk reference to natural source name
source_name = chunk.get('metadata', {}).get('source', 'unknown')
if source_name != 'unknown':
# Use just the filename without extension for natural reference
natural_name = Path(source_name).stem.replace('-', ' ').replace('_', ' ')
chunk_to_source[f'[chunk_{idx+1}]'] = f"the {natural_name} documentation"
else:
chunk_to_source[f'[chunk_{idx+1}]'] = "the documentation"
# Replace [chunk_X] with natural references instead of removing them
natural_answer = answer
for chunk_ref, natural_ref in chunk_to_source.items():
natural_answer = natural_answer.replace(chunk_ref, natural_ref)
# Clean up any remaining unreferenced citations (fallback)
natural_answer = re.sub(r'\[chunk_\d+\]', 'the documentation', natural_answer)
# Clean up multiple spaces and formatting
natural_answer = re.sub(r'\s+', ' ', natural_answer).strip()
return natural_answer, citations
def _calculate_confidence(self, answer: str, citations: List[Citation], chunks: List[Dict[str, Any]]) -> float:
"""
Calculate confidence score for the generated answer.
Args:
answer: Generated answer
citations: Extracted citations
chunks: Retrieved chunks
Returns:
Confidence score (0.0-1.0)
"""
if not chunks:
return 0.05 # No context = very low confidence
# Base confidence from context quality
scores = [chunk.get('score', 0) for chunk in chunks]
max_relevance = max(scores) if scores else 0
if max_relevance >= 0.8:
confidence = 0.7 # High-quality context
elif max_relevance >= 0.6:
confidence = 0.5 # Good context
elif max_relevance >= 0.4:
confidence = 0.3 # Fair context
else:
confidence = 0.1 # Poor context
# Uncertainty indicators
uncertainty_phrases = [
"does not contain sufficient information",
"context does not provide",
"insufficient information",
"cannot determine",
"not available in the provided documents"
]
if any(phrase in answer.lower() for phrase in uncertainty_phrases):
return min(0.15, confidence * 0.3)
# Citation bonus
if citations and chunks:
citation_ratio = len(citations) / min(len(chunks), 3)
confidence += 0.2 * citation_ratio
return min(confidence, 0.9) # Cap at 90%
def generate(
self,
query: str,
chunks: List[Dict[str, Any]]
) -> GeneratedAnswer:
"""
Generate an answer based on the query and retrieved chunks.
Args:
query: User's question
chunks: Retrieved document chunks
Returns:
GeneratedAnswer object with answer, citations, and metadata
"""
start_time = datetime.now()
# Check for no-context situation
if not chunks or all(len(chunk.get('content', chunk.get('text', ''))) < 20 for chunk in chunks):
return GeneratedAnswer(
answer="This information isn't available in the provided documents.",
citations=[],
confidence_score=0.05,
generation_time=0.1,
model_used=self.model_name,
context_used=chunks
)
# Format context from chunks
context = self._format_context(chunks)
# Create prompt using TechnicalPromptTemplates for consistency
prompt_data = TechnicalPromptTemplates.format_prompt_with_template(
query=query,
context=context
)
# Format for specific model types
if "squad" in self.model_name.lower() or "roberta" in self.model_name.lower():
# Squad2 uses special question/context format - handled in _call_api
prompt = f"Context: {context}\n\nQuestion: {query}"
elif "gpt2" in self.model_name.lower() or "distilgpt2" in self.model_name.lower():
# Simple completion style for GPT-2
prompt = f"""{prompt_data['system']}
{prompt_data['user']}
MANDATORY: Use [chunk_1], [chunk_2] etc. for all facts.
Answer:"""
elif "llama" in self.model_name.lower():
# Llama-2 chat format with technical templates
prompt = f"""[INST] {prompt_data['system']}
{prompt_data['user']}
MANDATORY: Use [chunk_1], [chunk_2] etc. for all facts. [/INST]"""
elif "mistral" in self.model_name.lower():
# Mistral instruction format with technical templates
prompt = f"""[INST] {prompt_data['system']}
{prompt_data['user']}
MANDATORY: Use [chunk_1], [chunk_2] etc. for all facts. [/INST]"""
elif "codellama" in self.model_name.lower():
# CodeLlama instruction format with technical templates
prompt = f"""[INST] {prompt_data['system']}
{prompt_data['user']}
MANDATORY: Use [chunk_1], [chunk_2] etc. for all facts. [/INST]"""
elif "distilbart" in self.model_name.lower():
# DistilBART is a summarization model - simpler prompt works better
prompt = f"""Technical Documentation Context:
{context}
Question: {query}
Instructions: Provide a technical answer using only the context above. Include source citations."""
else:
# Default instruction prompt with technical templates
prompt = f"""{prompt_data['system']}
{prompt_data['user']}
MANDATORY: Use [chunk_1], [chunk_2] etc. for all factual statements.
Answer:"""
# Generate response
try:
answer_with_citations = self._call_api(prompt)
# Extract and clean citations
clean_answer, citations = self._extract_citations(answer_with_citations, chunks)
# Calculate confidence
confidence = self._calculate_confidence(clean_answer, citations, chunks)
# Calculate generation time
generation_time = (datetime.now() - start_time).total_seconds()
return GeneratedAnswer(
answer=clean_answer,
citations=citations,
confidence_score=confidence,
generation_time=generation_time,
model_used=self.model_name,
context_used=chunks
)
except Exception as e:
logger.error(f"Error generating answer: {e}")
return GeneratedAnswer(
answer="I apologize, but I encountered an error while generating the answer. Please try again.",
citations=[],
confidence_score=0.0,
generation_time=0.0,
model_used=self.model_name,
context_used=chunks
)
def format_answer_with_citations(self, generated_answer: GeneratedAnswer) -> str:
"""
Format the generated answer with citations for display.
Args:
generated_answer: GeneratedAnswer object
Returns:
Formatted string with answer and citations
"""
formatted = f"{generated_answer.answer}\n\n"
if generated_answer.citations:
formatted += "**Sources:**\n"
for i, citation in enumerate(generated_answer.citations, 1):
formatted += f"{i}. {citation.source_file} (Page {citation.page_number})\n"
formatted += f"\n*Confidence: {generated_answer.confidence_score:.1%} | "
formatted += f"Model: {generated_answer.model_used} | "
formatted += f"Time: {generated_answer.generation_time:.2f}s*"
return formatted
if __name__ == "__main__":
# Example usage
generator = HuggingFaceAnswerGenerator()
# Example chunks (would come from retrieval system)
example_chunks = [
{
"id": "chunk_1",
"content": "RISC-V is an open-source instruction set architecture (ISA) based on reduced instruction set computer (RISC) principles.",
"metadata": {"page_number": 1, "source": "riscv-spec.pdf"},
"score": 0.95
}
]
# Generate answer
result = generator.generate(
query="What is RISC-V?",
chunks=example_chunks
)
# Display formatted result
print(generator.format_answer_with_citations(result))