Auto-Complete¶
The type system enables intelligent auto-complete suggestions during interactive graph building.
Overview¶
flowchart LR
STATE[Current State] --> ANALYZE[Analyze Output Type]
ANALYZE --> SUGGEST[Generate Suggestions]
SUGGEST --> RANK[Rank by Relevance]
RANK --> DISPLAY[Show to User]Context-Based Suggestions¶
Based on the last output type, the system suggests compatible operations:
from cbintel.chat import AutoComplete
complete = AutoComplete()
# Get suggestions based on current output type
suggestions = complete.suggest(last_output_type="Url[]")
for s in suggestions:
print(f"{s.operation}: {s.reason} (score: {s.score})")
Suggestion Mappings¶
After Url[]¶
- op: fetch_batch
score: 0.95
reason: "Fetch content from URLs"
- op: filter_urls
score: 0.70
reason: "Filter URLs by pattern"
- op: screenshot
score: 0.65
reason: "Capture screenshots"
After Html¶
- op: to_text
score: 0.90
reason: "Extract plain text"
- op: to_markdown
score: 0.85
reason: "Convert to markdown"
- op: extract_links
score: 0.70
reason: "Extract links from page"
After Html[]¶
- op: to_text_batch
score: 0.90
reason: "Extract text from all pages"
- op: to_markdown_batch
score: 0.85
reason: "Convert all to markdown"
After Text¶
- op: chunk
score: 0.90
reason: "Split into chunks for processing"
- op: entities
score: 0.85
reason: "Extract named entities"
- op: summarize
score: 0.80
reason: "Generate summary"
- op: embed
score: 0.75
reason: "Generate embedding"
After Chunk[]¶
- op: semantic_filter
score: 0.90
reason: "Filter relevant chunks"
- op: quality_filter
score: 0.85
reason: "Filter by quality"
- op: embed_batch
score: 0.80
reason: "Generate embeddings"
- op: integrate
score: 0.75
reason: "Synthesize into report"
After Entity[]¶
- op: filter_entities
score: 0.90
reason: "Filter by type/confidence"
- op: store_entities
score: 0.70
reason: "Store in knowledge base"
- op: to_report
score: 0.65
reason: "Format as report"
After Vector[]¶
- op: store_vectors
score: 0.90
reason: "Store in vector database"
- op: search_vectors
score: 0.80
reason: "Search similar content"
Suggestion Ranking¶
Suggestions are ranked by multiple factors:
Type Compatibility¶
Primary factor - operation must accept the current output type.
def type_score(output_type: str, operation: str) -> float:
expected_input = get_operation_input_type(operation)
if output_type == expected_input:
return 1.0
elif can_coerce(output_type, expected_input):
return 0.8
else:
return 0.0
Context Relevance¶
Based on what's already in the graph.
def context_score(graph: GraphDef, operation: str) -> float:
# If entities already extracted, don't suggest again
if has_operation(graph, "entities") and operation == "entities":
return 0.1
# If no synthesis yet, boost synthesis ops
if not has_synthesis(graph) and operation in ["integrate", "to_report"]:
return 1.2
return 1.0
Intent Alignment¶
Based on the original intent.
def intent_score(intent: str, operation: str) -> float:
# Person research should prioritize entity extraction
if intent == "person_research" and operation == "entities":
return 1.3
# Historical analysis should include archives
if intent == "historical_analysis" and operation == "fetch_archive":
return 1.3
return 1.0
Final Score¶
def rank_suggestion(suggestion: Suggestion, context: Context) -> float:
return (
suggestion.base_score *
type_score(context.output_type, suggestion.operation) *
context_score(context.graph, suggestion.operation) *
intent_score(context.intent, suggestion.operation)
)
Interactive Usage¶
In Chat Interface¶
Agent: Found 42 URLs. What would you like to do next?
Suggestions:
1. [Recommended] Fetch content from URLs
2. Filter URLs by domain
3. Take screenshots
4. Other...
User: 1
Agent: Fetching content from 42 URLs...
[Stage added: fetch_batch]
Agent: Retrieved 38 pages. Next?
Suggestions:
1. [Recommended] Extract plain text
2. Convert to markdown
3. Extract links
4. Other...
In Code¶
from cbintel.chat import InteractivePipelineBuilder
builder = InteractivePipelineBuilder()
# After each operation
suggestions = builder.get_suggestions()
# Display to user
for i, s in enumerate(suggestions):
print(f"{i+1}. {s.label}")
print(f" {s.description}")
# User selects
choice = int(input("Select: ")) - 1
builder.apply_suggestion(suggestions[choice])
Custom Suggestions¶
Register custom suggestion rules:
from cbintel.chat import register_suggestion_rule
@register_suggestion_rule(priority=10)
def suggest_geo_for_news(context: Context) -> list[Suggestion]:
"""Suggest geo routing for news research."""
if context.intent == "news_aggregation":
if not context.has_param("geo"):
return [Suggestion(
operation="set_param",
params={"geo": "us"},
score=0.8,
reason="Add geographic routing for regional news"
)]
return []
Filtering Suggestions¶
By Category¶
suggestions = complete.suggest(
last_output_type="Chunk[]",
categories=["filter", "synthesize"] # Only these categories
)
By Score Threshold¶
suggestions = complete.suggest(
last_output_type="Chunk[]",
min_score=0.7 # Only high-confidence suggestions
)
By Count¶
Suggestion Object¶
@dataclass
class Suggestion:
operation: str # Operation name
score: float # Relevance score (0-1)
reason: str # Human-readable reason
category: str # Operation category
params: dict # Suggested parameters
label: str # Display label
description: str # Detailed description
Configuration¶
Environment Variables¶
# Suggestion settings
AUTOCOMPLETE_MIN_SCORE=0.5
AUTOCOMPLETE_MAX_SUGGESTIONS=5
AUTOCOMPLETE_INCLUDE_DESCRIPTIONS=true
Customize Rankings¶
complete = AutoComplete(
min_score=0.6,
max_suggestions=5,
boost_synthesis=True, # Boost synthesis ops
penalize_redundant=True # Lower score for repeated ops
)
Best Practices¶
- Follow top suggestions - Usually most appropriate
- Consider context - Lower-ranked may be better for specific cases
- Check reasons - Understand why suggested
- Complete the pipeline - Don't stop without synthesis
- Review before execution - Validate the full graph