entari-plugin-hyw 4.0.0rc4__py3-none-any.whl → 4.0.0rc6__py3-none-any.whl

entari_plugin_hyw/modular_pipeline.py

@@ -0,0 +1,384 @@
+"""
+Modular Pipeline Dispatcher
+
+New pipeline architecture: Instruct Loop (x2) -> Summary.
+Simpler flow with self-correction/feedback loop.
+"""
+
+import asyncio
+import time
+from typing import Any, Dict, List, Optional, Callable, Awaitable
+
+from loguru import logger
+from openai import AsyncOpenAI
+
+from .stage_base import StageContext
+from .stage_instruct import InstructStage
+from .stage_instruct_deepsearch import InstructDeepsearchStage
+from .stage_summary import SummaryStage
+from .search import SearchService
+
+
+class ModularPipeline:
+    """
+    Modular Pipeline.
+    
+    Flow:
+    1. Instruct: Initial Discovery + Mode Decision (fast/deepsearch).
+    2. [Deepsearch only] Instruct Deepsearch Loop: Supplement info (max 3 iterations).
+    3. Summary: Generate final response.
+    """
+    
+    def __init__(self, config: Any, send_func: Optional[Callable[[str], Awaitable[None]]] = None):
+        self.config = config
+        self.send_func = send_func
+        self.search_service = SearchService(config)
+        self.client = AsyncOpenAI(base_url=config.base_url, api_key=config.api_key)
+        
+        # Initialize stages
+        self.instruct_stage = InstructStage(config, self.search_service, self.client)
+        self.instruct_deepsearch_stage = InstructDeepsearchStage(config, self.search_service, self.client)
+        self.summary_stage = SummaryStage(config, self.search_service, self.client)
+    
+    async def execute(
+        self,
+        user_input: str,
+        conversation_history: List[Dict],
+        model_name: str = None,
+        images: List[str] = None,
+        vision_model_name: str = None,
+        selected_vision_model: str = None,
+    ) -> Dict[str, Any]:
+        """Execute the modular pipeline."""
+        start_time = time.time()
+        stats = {"start_time": start_time}
+        usage_totals = {"input_tokens": 0, "output_tokens": 0}
+        active_model = model_name or self.config.model_name
+        
+        context = StageContext(
+            user_input=user_input,
+            images=images or [],
+            conversation_history=conversation_history,
+        )
+        
+        # Determine if model supports image input
+        model_cfg_dict = next((m for m in self.config.models if m.get("name") == active_model), None)
+        if model_cfg_dict:
+            context.image_input_supported = model_cfg_dict.get("image_input", True)
+        else:
+             context.image_input_supported = True # Default to True if unknown
+             
+        logger.info(f"Pipeline Execution: Model '{active_model}' Image Input Supported: {context.image_input_supported}")
+
+        
+        trace: Dict[str, Any] = {
+            "instruct_rounds": [],
+            "summary": None,
+        }
+        
+        try:
+            logger.info(f"Pipeline: Processing '{user_input[:30]}...'")
+            
+            # === Stage 1: Instruct (Initial Discovery) ===
+            logger.info("Pipeline: Stage 1 - Instruct")
+            instruct_result = await self.instruct_stage.execute(context)
+            
+            # Trace & Usage
+            instruct_result.trace["stage_name"] = "Instruct (Round 1)"
+            trace["instruct_rounds"].append(instruct_result.trace)
+            usage_totals["input_tokens"] += instruct_result.usage.get("input_tokens", 0)
+            usage_totals["output_tokens"] += instruct_result.usage.get("output_tokens", 0)
+            
+            # Check refuse
+            if context.should_refuse:
+                return self._build_refusal_response(context, conversation_history, active_model, stats)
+
+            # === Stage 2: Deepsearch Loop (if mode is deepsearch) ===
+            if context.selected_mode == "deepsearch":
+                MAX_DEEPSEARCH_ITERATIONS = 3
+                logger.info(f"Pipeline: Mode is 'deepsearch', starting loop (max {MAX_DEEPSEARCH_ITERATIONS} iterations)")
+                
+                for i in range(MAX_DEEPSEARCH_ITERATIONS):
+                    logger.info(f"Pipeline: Stage 2 - Deepsearch Iteration {i + 1}")
+                    deepsearch_result = await self.instruct_deepsearch_stage.execute(context)
+                    
+                    # Trace & Usage
+                    deepsearch_result.trace["stage_name"] = f"Deepsearch (Iteration {i + 1})"
+                    trace["instruct_rounds"].append(deepsearch_result.trace)
+                    usage_totals["input_tokens"] += deepsearch_result.usage.get("input_tokens", 0)
+                    usage_totals["output_tokens"] += deepsearch_result.usage.get("output_tokens", 0)
+                    
+                    # Check if should stop
+                    if deepsearch_result.data.get("should_stop"):
+                        logger.info(f"Pipeline: Deepsearch loop ended at iteration {i + 1}")
+                        break
+            else:
+                logger.info("Pipeline: Mode is 'fast', skipping deepsearch stage")
+            
+            # === Stage 3: Summary ===
+            # Collect page screenshots if image mode (already rendered in InstructStage)
+            all_images = list(images) if images else []
+            
+            if context.image_input_supported:
+                # Collect pre-rendered screenshots from web_results
+                for r in context.web_results:
+                    if r.get("_type") == "page" and r.get("screenshot_b64"):
+                        all_images.append(r["screenshot_b64"])
+            
+            summary_result = await self.summary_stage.execute(
+                context,
+                images=all_images if all_images else None
+            )
+            trace["summary"] = summary_result.trace
+            usage_totals["input_tokens"] += summary_result.usage.get("input_tokens", 0)
+            usage_totals["output_tokens"] += summary_result.usage.get("output_tokens", 0)
+            
+            summary_content = summary_result.data.get("content", "")
+            
+            # === Result Assembly ===
+            stats["total_time"] = time.time() - start_time
+            structured = self._parse_response(summary_content, context)
+            
+            # === Image Caching (Prefetch images for UI) ===
+            try:
+                from .image_cache import get_image_cache
+                cache = get_image_cache()
+                
+                # 1. Collect all image URLs from structured response
+                all_image_urls = []
+                for ref in structured.get("references", []):
+                    if ref.get("images"):
+                        all_image_urls.extend([img for img in ref["images"] if img and img.startswith("http")])
+                
+                if all_image_urls:
+                    # 2. Prefetch (wait for them as we are about to render)
+                    cached_map = await cache.get_all_cached(all_image_urls)
+                    
+                    # 3. Update structured response with cached (base64) URLs
+                    for ref in structured.get("references", []):
+                        if ref.get("images"):
+                            # Keep cached images, but preserve original URLs as fallback
+                            new_images = []
+                            for img in ref["images"]:
+                                # 1. Already Base64 (from Search Injection) -> Keep it
+                                if img.startswith("data:"):
+                                    new_images.append(img)
+                                    continue
+
+                                # 2. Cached successfully -> Keep it
+                                cached_val = cached_map.get(img)
+                                if cached_val and cached_val.startswith("data:"):
+                                    new_images.append(cached_val)
+                                # 3. Else -> DROP IT (User request: "Delete Fallback, must download in advance")
+                            ref["images"] = new_images
+            except Exception as e:
+                logger.warning(f"Pipeline: Image caching failed: {e}")
+            
+            # Debug: Log image counts
+            total_ref_images = sum(len(ref.get("images", []) or []) for ref in structured.get("references", []))
+            logger.info(f"Pipeline: Final structured response has {len(structured.get('references', []))} refs with {total_ref_images} images total")
+            
+            stages_used = self._build_stages_ui(trace, context, images)
+            
+            conversation_history.append({"role": "user", "content": user_input})
+            conversation_history.append({"role": "assistant", "content": summary_content})
+            
+            return {
+                "llm_response": summary_content,
+                "structured_response": structured,
+                "stats": stats,
+                "model_used": active_model,
+                "conversation_history": conversation_history,
+                "trace_markdown": self._render_trace_markdown(trace),
+                "billing_info": {
+                    "input_tokens": usage_totals["input_tokens"],
+                    "output_tokens": usage_totals["output_tokens"],
+                    "total_cost": 0.0
+                },
+                "stages_used": stages_used,
+                "web_results": context.web_results,
+            }
+
+        except Exception as e:
+            logger.error(f"Pipeline: Critical Error - {e}")
+            import traceback
+            logger.error(traceback.format_exc())
+            return {
+                "llm_response": f"Error: {e}",
+                "stats": stats,
+                "error": str(e)
+            }
+
+    def _build_refusal_response(self, context, history, model, stats):
+        return {
+            "llm_response": "Refused",
+            "structured_response": {},
+            "stats": stats,
+            "model_used": model,
+            "conversation_history": history,
+            "refuse_answer": True,
+            "refuse_reason": context.refuse_reason
+        }
+
+    def _parse_response(self, text: str, context: StageContext) -> Dict[str, Any]:
+        """Parse response and extract citations, prioritizing fetched items."""
+        import re
+        parsed = {"response": "", "references": [], "page_references": [], "image_references": []}
+        if not text: return parsed
+        
+        # Simple cleanup
+        ref_pattern = re.compile(r'(?:\n\s*|^)\s*(?:#{1,3}|\*\*)\s*(?:References|Citations|Sources|参考资料)[\s\S]*$', re.IGNORECASE | re.MULTILINE)
+        body_text = ref_pattern.sub('', text)
+        
+        # 1. Identify all cited numeric IDs from [N]
+        cited_ids = []
+        for m in re.finditer(r'\[(\d+)\]', body_text):
+            try:
+                cid = int(m.group(1))
+                if cid not in cited_ids: cited_ids.append(cid)
+            except: pass
+            
+        # 2. Collect cited items and determine "is_fetched" status
+        cited_items = []
+        for cid in cited_ids:
+            item = next((r for r in context.web_results if r.get("_id") == cid), None)
+            if not item: continue
+            
+            # Check if this URL was fetched (appears as a "page" result)
+            is_fetched = any(r.get("_type") == "page" and r.get("url") == item.get("url") for r in context.web_results)
+            cited_items.append({
+                "original_id": cid,
+                "item": item,
+                "is_fetched": is_fetched
+            })
+            
+        # 3. Sort: Fetched pages first, then regular search results
+        cited_items.sort(key=lambda x: x["is_fetched"], reverse=True)
+        
+        # 4. Create Re-indexing Map
+        reindex_map = {}
+        for i, entry in enumerate(cited_items):
+            reindex_map[entry["original_id"]] = i + 1
+            
+            # Populate result references in sorted order
+            item = entry["item"]
+            ref_entry = {
+                "title": item.get("title", ""),
+                "url": item.get("url", ""),
+                "domain": item.get("domain", ""),
+                "snippet": (item.get("content", "") or "")[:200] + "...", # More snippet
+                "is_fetched": entry["is_fetched"],
+                "type": item.get("_type", "search"),
+                "raw_screenshot_b64": item.get("raw_screenshot_b64"),  # Real page screenshot for Sources
+                "images": item.get("images"),
+            }
+            # Add to unified list (frontend can handle splitting if needed, but we provide sorted order)
+            parsed["references"].append(ref_entry)
+            
+        # 5. Replace [N] in text with new indices
+        def repl(m):
+            try:
+                oid = int(m.group(1))
+                return f"[{reindex_map[oid]}]" if oid in reindex_map else m.group(0)
+            except: return m.group(0)
+            
+        parsed["response"] = re.sub(r'\[(\d+)\]', repl, body_text).strip()
+        return parsed
+
+    def _build_stages_ui(self, trace: Dict[str, Any], context: StageContext, images: List[str]) -> List[Dict[str, Any]]:
+        stages = []
+        
+        # 1. Search Results
+        search_refs = []
+        seen = set()
+        for r in context.web_results:
+            if r.get("_type") == "search" and r.get("url") not in seen:
+                seen.add(r["url"])
+                is_fetched = any(p.get("url") == r["url"] for p in context.web_results if p.get("_type") == "page")
+                search_refs.append({
+                    "title": r.get("title", ""),
+                    "url": r["url"],
+                    "snippet": (r.get("content", "") or "")[:100] + "...",
+                    "is_fetched": is_fetched
+                })
+        
+        # Sort: Fetched first
+        search_refs.sort(key=lambda x: x["is_fetched"], reverse=True)
+        
+        if search_refs:
+            stages.append({
+                "name": "Search",
+                "model": "Web Search",
+                "icon_config": "openai",
+                "provider": "Web",
+                "references": search_refs,
+                "description": f"Found {len(search_refs)} results."
+            })
+            
+        # 2. Instruct Rounds
+        for i, t in enumerate(trace.get("instruct_rounds", [])):
+            stage_name = t.get("stage_name", f"Analysis {i+1}")
+            tool_count = t.get("tool_calls", 0)
+            desc = t.get("output", "")
+            
+            if tool_count > 0:
+                # If tools were used, prefer showing tool info even if there's reasoning
+                desc = f"Executed {tool_count} tool calls."
+            elif not desc:
+                desc = "Processing..."
+            
+            # Calculate cost from config prices
+            usage = t.get("usage", {})
+            instruct_cfg = self.config.get_model_config("instruct")
+            input_price = instruct_cfg.get("input_price") or 0
+            output_price = instruct_cfg.get("output_price") or 0
+            cost = (usage.get("input_tokens", 0) * input_price + usage.get("output_tokens", 0) * output_price) / 1_000_000
+            
+            stages.append({
+                "name": stage_name,
+                "model": t.get("model"),
+                "icon_config": "google",
+                "provider": "Instruct",
+                "time": t.get("time", 0),
+                "description": desc,
+                "usage": usage,
+                "cost": cost
+            })
+            
+        # 3. Summary
+        if trace.get("summary"):
+            s = trace["summary"]
+            usage = s.get("usage", {})
+            main_cfg = self.config.get_model_config("main")
+            input_price = main_cfg.get("input_price") or 0
+            output_price = main_cfg.get("output_price") or 0
+            cost = (usage.get("input_tokens", 0) * input_price + usage.get("output_tokens", 0) * output_price) / 1_000_000
+            
+            stages.append({
+                "name": "Summary",
+                "model": s.get("model"),
+                "icon_config": "google",
+                "provider": "Summary",
+                "time": s.get("time", 0),
+                "description": "Generated final answer.",
+                "usage": usage,
+                "cost": cost
+            })
+            
+        return stages
+
+    def _render_trace_markdown(self, trace: Dict[str, Any]) -> str:
+        parts = ["# Pipeline Trace\n"]
+        if trace.get("instruct_rounds"):
+            parts.append(f"## Instruct ({len(trace['instruct_rounds'])} rounds)\n")
+            for i, r in enumerate(trace["instruct_rounds"]):
+                 name = r.get("stage_name", f"Round {i+1}")
+                 parts.append(f"### {name}\n" + str(r))
+        if trace.get("summary"):
+            parts.append("## Summary\n" + str(trace["summary"]))
+        return "\n".join(parts)
+    
+    async def close(self):
+        try:
+            await self.search_service.close()
+        except: pass