llm_council/backend/council.py
Irina Levit 3546c04348
Some checks failed
CI / backend-test (push) Successful in 4m9s
CI / frontend-test (push) Failing after 3m48s
CI / lint-python (push) Successful in 1m41s
CI / secret-scanning (push) Successful in 1m20s
CI / dependency-scan (push) Successful in 10m50s
CI / workflow-summary (push) Successful in 1m11s
feat: Major UI/UX improvements and production readiness
## Features Added

### Document Reference System
- Implemented numbered document references (@1, @2, etc.) with autocomplete dropdown
- Added fuzzy filename matching for @filename references
- Document filtering now prioritizes numeric refs > filename refs > all documents
- Autocomplete dropdown appears when typing @ with keyboard navigation (Up/Down, Enter/Tab, Escape)
- Document numbers displayed in UI for easy reference

### Conversation Management
- Added conversation rename functionality with inline editing
- Implemented conversation search (by title and content)
- Search box always visible, even when no conversations exist
- Export reports now replace @N references with actual filenames

### UI/UX Improvements
- Removed debug toggle button
- Improved text contrast in dark mode (better visibility)
- Made input textarea expand to full available width
- Fixed file text color for better readability
- Enhanced document display with numbered badges

### Configuration & Timeouts
- Made HTTP client timeouts configurable (connect, write, pool)
- Added .env.example with all configuration options
- Updated timeout documentation

### Developer Experience
- Added `make test-setup` target for automated test conversation creation
- Test setup script supports TEST_MESSAGE and TEST_DOCS env vars
- Improved Makefile with dev and test-setup targets

### Documentation
- Updated ARCHITECTURE.md with all new features
- Created comprehensive deployment documentation
- Added GPU VM setup guides
- Removed unnecessary markdown files (CLAUDE.md, CONTRIBUTING.md, header.jpg)
- Organized documentation in docs/ directory

### GPU VM / Ollama (Stability + GPU Offload)
- Updated GPU VM docs to reflect the working systemd environment for remote Ollama
- Standardized remote Ollama port to 11434 (and added /v1/models verification)
- Documented required env for GPU offload on this VM:
  - `OLLAMA_MODELS=/mnt/data/ollama`, `HOME=/mnt/data/ollama/home`
  - `OLLAMA_LLM_LIBRARY=cuda_v12` (not `cuda`)
  - `LD_LIBRARY_PATH=/usr/local/lib/ollama:/usr/local/lib/ollama/cuda_v12`

## Technical Changes

### Backend
- Enhanced `docs_context.py` with reference parsing (numeric and filename)
- Added `update_conversation_title` to storage.py
- New endpoints: PATCH /api/conversations/{id}/title, GET /api/conversations/search
- Improved report generation with filename substitution

### Frontend
- Removed debugMode state and related code
- Added autocomplete dropdown component
- Implemented search functionality in Sidebar
- Enhanced ChatInterface with autocomplete and improved textarea sizing
- Updated CSS for better contrast and responsive design

## Files Changed
- Backend: config.py, council.py, docs_context.py, main.py, storage.py
- Frontend: App.jsx, ChatInterface.jsx, Sidebar.jsx, and related CSS files
- Documentation: README.md, ARCHITECTURE.md, new docs/ directory
- Configuration: .env.example, Makefile
- Scripts: scripts/test_setup.py

## Breaking Changes
None - all changes are backward compatible

## Testing
- All existing tests pass
- New test-setup script validates conversation creation workflow
- Manual testing of autocomplete, search, and rename features
2025-12-28 18:15:02 -05:00

538 lines
19 KiB
Python

"""3-stage LLM Council orchestration."""
import time
from typing import List, Dict, Any, Tuple, Optional
from .llm_client import query_models_parallel, query_model
from .config import COUNCIL_MODELS, CHAIRMAN_MODEL, CHAIRMAN_TIMEOUT_SECONDS, TITLE_GENERATION_TIMEOUT_SECONDS
def _format_docs_context(docs_text: Optional[str]) -> str:
if not docs_text or not docs_text.strip():
return ""
return (
"\n\nREFERENCE DOCUMENTS (user-provided markdown):\n"
"Use these as additional context if relevant. Quote sparingly and cite sections when helpful.\n"
f"{docs_text.strip()}\n"
)
async def stage1_collect_responses(user_query: str, docs_text: Optional[str] = None) -> Tuple[List[Dict[str, Any]], Dict[str, Any]]:
"""
Stage 1: Collect individual responses from all council models.
Args:
user_query: The user's question
Returns:
Tuple of (results list, metadata dict with timing info)
"""
start_time = time.time()
prompt = f"{user_query}{_format_docs_context(docs_text)}"
messages = [{"role": "user", "content": prompt}]
# Query all models in parallel
responses = await query_models_parallel(COUNCIL_MODELS, messages)
duration = time.time() - start_time
# Format results
stage1_results = []
successful_models = []
failed_models = []
for model, response in responses.items():
if response is not None: # Only include successful responses
stage1_results.append({
"model": model,
"response": response.get('content', '')
})
successful_models.append(model)
else:
failed_models.append(model)
metadata = {
"duration_seconds": round(duration, 2),
"successful_models": successful_models,
"failed_models": failed_models,
"total_models": len(COUNCIL_MODELS)
}
return stage1_results, metadata
async def stage1_collect_responses_streaming(
user_query: str,
docs_text: Optional[str] = None,
on_response = None
) -> Tuple[List[Dict[str, Any]], Dict[str, Any]]:
"""
Stage 1: Collect individual responses from all council models, streaming as they complete.
Args:
user_query: The user's question
docs_text: Optional document context
on_response: Optional callback(model, response_dict) called as each response completes
Returns:
Tuple of (results list, metadata dict with timing info)
"""
import asyncio
from .llm_client import query_model, LLM_TIMEOUT_SECONDS, MAX_TOKENS
start_time = time.time()
prompt = f"{user_query}{_format_docs_context(docs_text)}"
messages = [{"role": "user", "content": prompt}]
# Query all models in parallel, but yield results as they complete
async def query_and_notify(model: str):
response = await query_model(model, messages, timeout=LLM_TIMEOUT_SECONDS, max_tokens_override=MAX_TOKENS)
if on_response and response is not None:
result = {"model": model, "response": response.get('content', '')}
await on_response(model, result)
return model, response
tasks = [query_and_notify(model) for model in COUNCIL_MODELS]
responses_dict = {}
# Use as_completed to process results as they finish
for coro in asyncio.as_completed(tasks):
model, response = await coro
responses_dict[model] = response
duration = time.time() - start_time
# Format results
stage1_results = []
successful_models = []
failed_models = []
for model, response in responses_dict.items():
if response is not None: # Only include successful responses
stage1_results.append({
"model": model,
"response": response.get('content', '')
})
successful_models.append(model)
else:
failed_models.append(model)
metadata = {
"duration_seconds": round(duration, 2),
"successful_models": successful_models,
"failed_models": failed_models,
"total_models": len(COUNCIL_MODELS)
}
return stage1_results, metadata
async def stage2_collect_rankings(
user_query: str,
stage1_results: List[Dict[str, Any]],
docs_text: Optional[str] = None,
) -> Tuple[List[Dict[str, Any]], Dict[str, str], Dict[str, Any]]:
"""
Stage 2: Each model ranks the anonymized responses.
Args:
user_query: The original user query
stage1_results: Results from Stage 1
Returns:
Tuple of (rankings list, label_to_model mapping, metadata dict with timing)
"""
start_time = time.time()
# Handle empty stage1_results
if not stage1_results:
return [], {}, {"duration_seconds": 0.0, "successful_models": [], "failed_models": [], "total_models": len(COUNCIL_MODELS)}
# Create anonymized labels for responses (Response A, Response B, etc.)
labels = [chr(65 + i) for i in range(len(stage1_results))] # A, B, C, ...
# Create mapping from label to model name
label_to_model = {
f"Response {label}": result['model']
for label, result in zip(labels, stage1_results)
}
# Build the ranking prompt
responses_text = "\n\n".join([
f"Response {label}:\n{result['response']}"
for label, result in zip(labels, stage1_results)
])
ranking_prompt = f"""You are evaluating different responses to the following question:
Question: {user_query}
{_format_docs_context(docs_text)}
Here are the responses from different models (anonymized):
{responses_text}
Your task:
1. First, evaluate each response individually. For each response, explain what it does well and what it does poorly.
2. Then, at the very end of your response, provide a final ranking.
IMPORTANT: Your final ranking MUST be formatted EXACTLY as follows:
- Start with the line "FINAL RANKING:" (all caps, with colon)
- Then list the responses from best to worst as a numbered list
- Each line should be: number, period, space, then ONLY the response label (e.g., "1. Response A")
- Do not add any other text or explanations in the ranking section
Example of the correct format for your ENTIRE response:
Response A provides good detail on X but misses Y...
Response B is accurate but lacks depth on Z...
Response C offers the most comprehensive answer...
FINAL RANKING:
1. Response C
2. Response A
3. Response B
Now provide your evaluation and ranking:"""
messages = [{"role": "user", "content": ranking_prompt}]
# Get rankings from all council models in parallel
responses = await query_models_parallel(COUNCIL_MODELS, messages)
duration = time.time() - start_time
# Format results
stage2_results = []
successful_models = []
failed_models = []
for model, response in responses.items():
if response is not None:
full_text = response.get('content', '')
parsed = parse_ranking_from_text(full_text)
stage2_results.append({
"model": model,
"ranking": full_text,
"parsed_ranking": parsed
})
successful_models.append(model)
else:
failed_models.append(model)
metadata = {
"duration_seconds": round(duration, 2),
"successful_models": successful_models,
"failed_models": failed_models,
"total_models": len(COUNCIL_MODELS)
}
return stage2_results, label_to_model, metadata
async def stage3_synthesize_final(
user_query: str,
stage1_results: List[Dict[str, Any]],
stage2_results: List[Dict[str, Any]],
docs_text: Optional[str] = None,
) -> Tuple[Dict[str, Any], Dict[str, Any]]:
"""
Stage 3: Chairman synthesizes final response.
Args:
user_query: The original user query
stage1_results: Individual model responses from Stage 1
stage2_results: Rankings from Stage 2
Returns:
Tuple of (result dict with 'model' and 'response' keys, metadata dict with timing)
"""
start_time = time.time()
# Handle empty inputs
if not stage1_results:
duration = time.time() - start_time
return {
"model": CHAIRMAN_MODEL,
"response": "Error: Cannot synthesize final answer - no responses from Stage 1."
}, {
"duration_seconds": round(duration, 2),
"model": CHAIRMAN_MODEL,
"success": False
}
# Build comprehensive context for chairman
# Truncate very long responses to avoid exceeding token/context limits
# More aggressive truncation to keep total prompt under ~2000 tokens (~8000 chars)
MAX_RESPONSE_LENGTH = 2000 # Characters per response
MAX_RANKING_LENGTH = 1000 # Characters per ranking
MAX_DOCS_LENGTH = 2000 # Max characters for docs context
MAX_TOTAL_PROMPT_LENGTH = 8000 # Max total prompt length (safety limit)
def truncate_text(text: str, max_length: int) -> str:
"""Truncate text to max_length, adding ellipsis if truncated."""
if len(text) <= max_length:
return text
return text[:max_length-3] + "..."
# Truncate docs_text if provided
truncated_docs_text = docs_text
if docs_text and len(docs_text) > MAX_DOCS_LENGTH:
truncated_docs_text = truncate_text(docs_text, MAX_DOCS_LENGTH)
stage1_text = "\n\n".join([
f"Model: {result['model']}\nResponse: {truncate_text(result['response'], MAX_RESPONSE_LENGTH)}"
for result in stage1_results
])
stage2_text = "\n\n".join([
f"Model: {result['model']}\nRanking: {truncate_text(result['ranking'], MAX_RANKING_LENGTH)}"
for result in stage2_results
]) if stage2_results else "No rankings available from Stage 2."
chairman_prompt = f"""You are the Chairman of an LLM Council. Multiple AI models have provided responses to a user's question, and then ranked each other's responses.
Original Question: {user_query}
{_format_docs_context(truncated_docs_text)}
STAGE 1 - Individual Responses:
{stage1_text}
STAGE 2 - Peer Rankings:
{stage2_text}
Your task as Chairman is to synthesize all of this information into a single, comprehensive, accurate answer to the user's original question. Consider:
- The individual responses and their insights
- The peer rankings and what they reveal about response quality
- Any patterns of agreement or disagreement
Provide a clear, well-reasoned final answer that represents the council's collective wisdom:"""
# Apply final safety truncation if prompt is still too long
if len(chairman_prompt) > MAX_TOTAL_PROMPT_LENGTH:
# Truncate the prompt itself if it exceeds the limit
chairman_prompt = chairman_prompt[:MAX_TOTAL_PROMPT_LENGTH - 100] + "\n\n[Content truncated due to length limits...]\n\nProvide a clear, well-reasoned final answer:"
messages = [{"role": "user", "content": chairman_prompt}]
# Query the chairman model
# Note: For very long prompts, we might need to truncate or summarize
# For now, we'll try with the full prompt and handle errors gracefully
# Use the default max_tokens (2048) to stay within credit limits
# If you have more credits, you can increase MAX_TOKENS in config.py
response = await query_model(CHAIRMAN_MODEL, messages, timeout=CHAIRMAN_TIMEOUT_SECONDS)
duration = time.time() - start_time
if response is None:
# Try to get more specific error info - check if prompt might be too long
prompt_length = len(chairman_prompt)
estimated_tokens = prompt_length // 4 # Rough estimate: ~4 chars per token
error_msg = (
"Error: Unable to generate final synthesis.\n\n"
"The chairman model failed to respond. Possible causes:\n"
"- Model '{}' not available on the server\n"
"- Server not running or unreachable\n"
"- Network/API errors\n"
"- Prompt too long (estimated ~{} tokens)\n"
"- Server timeout or overloaded\n\n"
"Check the backend terminal logs for the exact error message."
).format(CHAIRMAN_MODEL, estimated_tokens)
return {
"model": CHAIRMAN_MODEL,
"response": error_msg
}, {
"duration_seconds": round(duration, 2),
"model": CHAIRMAN_MODEL,
"success": False
}
return {
"model": CHAIRMAN_MODEL,
"response": response.get('content', '')
}, {
"duration_seconds": round(duration, 2),
"model": CHAIRMAN_MODEL,
"success": True
}
def parse_ranking_from_text(ranking_text: str) -> List[str]:
"""
Parse the FINAL RANKING section from the model's response.
Args:
ranking_text: The full text response from the model
Returns:
List of response labels in ranked order
"""
import re
# Look for "FINAL RANKING:" section
if "FINAL RANKING:" in ranking_text:
# Extract everything after "FINAL RANKING:"
parts = ranking_text.split("FINAL RANKING:")
if len(parts) >= 2:
ranking_section = parts[1]
# Try to extract numbered list format (e.g., "1. Response A")
# This pattern looks for: number, period, optional space, "Response X"
numbered_matches = re.findall(r'\d+\.\s*Response [A-Z]', ranking_section)
if numbered_matches:
# Extract just the "Response X" part
return [re.search(r'Response [A-Z]', m).group() for m in numbered_matches]
# Fallback: Extract all "Response X" patterns in order
matches = re.findall(r'Response [A-Z]', ranking_section)
return matches
# Fallback: try to find any "Response X" patterns in order
matches = re.findall(r'Response [A-Z]', ranking_text)
return matches
def calculate_aggregate_rankings(
stage2_results: List[Dict[str, Any]],
label_to_model: Dict[str, str]
) -> List[Dict[str, Any]]:
"""
Calculate aggregate rankings across all models.
Args:
stage2_results: Rankings from each model
label_to_model: Mapping from anonymous labels to model names
Returns:
List of dicts with model name and average rank, sorted best to worst
"""
from collections import defaultdict
# Track positions for each model
model_positions = defaultdict(list)
for ranking in stage2_results:
ranking_text = ranking['ranking']
# Parse the ranking from the structured format
parsed_ranking = parse_ranking_from_text(ranking_text)
for position, label in enumerate(parsed_ranking, start=1):
if label in label_to_model:
model_name = label_to_model[label]
model_positions[model_name].append(position)
# Calculate average position for each model
aggregate = []
for model, positions in model_positions.items():
if positions:
avg_rank = sum(positions) / len(positions)
aggregate.append({
"model": model,
"average_rank": round(avg_rank, 2),
"rankings_count": len(positions)
})
# Sort by average rank (lower is better)
aggregate.sort(key=lambda x: x['average_rank'])
return aggregate
async def generate_conversation_title(user_query: str) -> str:
"""
Generate a short title for a conversation based on the first user message.
Args:
user_query: The first user message
Returns:
A short title (3-5 words)
"""
title_prompt = f"""Generate a very short title (3-5 words maximum) that summarizes the following question.
The title should be concise and descriptive. Do not use quotes or punctuation in the title.
Question: {user_query}
Title:"""
messages = [{"role": "user", "content": title_prompt}]
# Use chairman model for title generation
# Use configurable timeout (may need longer for local models which load on first request)
response = await query_model(CHAIRMAN_MODEL, messages, timeout=TITLE_GENERATION_TIMEOUT_SECONDS, max_tokens_override=50)
if response is None:
# Fallback to a generic title
return "New Conversation"
title = response.get('content', 'New Conversation').strip()
# Clean up the title - remove quotes, limit length
title = title.strip('"\'')
# Truncate if too long
if len(title) > 50:
title = title[:47] + "..."
return title
async def run_full_council(user_query: str, docs_text: Optional[str] = None) -> Tuple[List, List, Dict, Dict]:
"""
Run the complete 3-stage council process.
Args:
user_query: The user's question
Returns:
Tuple of (stage1_results, stage2_results, stage3_result, metadata)
"""
total_start_time = time.time()
# Stage 1: Collect individual responses
stage1_results, stage1_metadata = await stage1_collect_responses(user_query, docs_text=docs_text)
# If no models responded successfully, return error with helpful message
if not stage1_results:
error_msg = (
"All models failed to respond. This could be due to:\n"
"- Server not running or unreachable\n"
"- Model names not available on the server\n"
"- Network/API errors\n"
"- Server timeout or overloaded\n"
"- Invalid OPENAI_COMPAT_BASE_URL configuration\n\n"
"Check the backend logs for detailed error messages."
)
total_duration = time.time() - total_start_time
return [], [], {
"model": "error",
"response": error_msg
}, {
"label_to_model": {},
"aggregate_rankings": {},
"stage1_metadata": stage1_metadata,
"stage2_metadata": {},
"stage3_metadata": {},
"total_duration_seconds": round(total_duration, 2)
}
# Stage 2: Collect rankings
stage2_results, label_to_model, stage2_metadata = await stage2_collect_rankings(user_query, stage1_results, docs_text=docs_text)
# Calculate aggregate rankings
aggregate_rankings = calculate_aggregate_rankings(stage2_results, label_to_model)
# Stage 3: Synthesize final answer
stage3_result, stage3_metadata = await stage3_synthesize_final(
user_query,
stage1_results,
stage2_results,
docs_text=docs_text,
)
total_duration = time.time() - total_start_time
# Prepare metadata
metadata = {
"label_to_model": label_to_model,
"aggregate_rankings": aggregate_rankings,
"stage1_metadata": stage1_metadata,
"stage2_metadata": stage2_metadata,
"stage3_metadata": stage3_metadata,
"total_duration_seconds": round(total_duration, 2)
}
return stage1_results, stage2_results, stage3_result, metadata