Sadum1's picture
Upload folder using huggingface_hub
7376e73 verified
Raw History Blame Contribute Delete
6.82 kB
"""LangGraph orchestrator for V2 + Reflexion pipeline (Phase 5 + 6).
Pipeline:
QueryExpander → Searcher → QualityCritic → Summarizer
↓
Synthesizer ←──────────┐
↓ │ needs_revision=True
Critic ─────────────┘ (max 3 iterations)
↓ needs_revision=False
TrendAnalyst → END
Pédagogique — LangGraph conditional edges:
`add_conditional_edges(source, routing_fn, mapping)` adds a branch after
`source`. `routing_fn(state) -> str` returns the name of the next node.
`mapping` is a dict {return_value: node_name} used by LangGraph to
validate that all branches are declared.
Public API:
build_graph() -> compiled LangGraph
run_pipeline(topic, **config_kwargs) -> final WatchState dict
"""
from __future__ import annotations
import logging
from typing import Any
from langgraph.graph import END, StateGraph
from src.agents import (
critic,
quality_critic,
query_expander,
searcher,
summarizer,
synthesizer,
trend_analyst,
)
from src.agents.state import RunConfig, WatchState
from src.config import DEFAULT_DOMAIN_HINT, DEFAULT_FROM_YEAR, MAX_REFLEXION_ITERATIONS
logger = logging.getLogger(__name__)
# ============================================================
# Node names — constants to avoid typos in edges
# ============================================================
N_EXPANDER = "query_expander"
N_SEARCHER = "searcher"
N_CRITIC_Q = "quality_critic"
N_SUMMARIZER = "summarizer"
N_SYNTHESIZER = "synthesizer"
N_CRITIC = "critic"
N_TRENDS = "trend_analyst"
# ============================================================
# Routing function — the heart of the Reflexion loop
# ============================================================
def route_after_critic(state: WatchState) -> str:
"""Decide what comes after the Critic.
Returns N_SYNTHESIZER if the synthesis needs revision and we haven't
hit the iteration cap. Returns N_TRENDS otherwise.
This function is called by LangGraph after every Critic run.
"""
feedbacks = state.get("critic_feedbacks", [])
iteration = state.get("synthesis_iteration", 0)
if not feedbacks:
return N_TRENDS
last = feedbacks[-1]
if last.needs_revision and iteration < MAX_REFLEXION_ITERATIONS:
logger.info(
"[Router] iteration=%d quality=%s → looping back to Synthesizer",
iteration, last.overall_quality,
)
return N_SYNTHESIZER
logger.info(
"[Router] iteration=%d quality=%s → proceeding to TrendAnalyst",
iteration, last.overall_quality,
)
return N_TRENDS
# ============================================================
# Graph builder
# ============================================================
def build_graph():
"""Wire all agents into a LangGraph with Reflexion loop and compile it."""
g: StateGraph = StateGraph(WatchState)
# Register every node (name → callable).
g.add_node(N_EXPANDER, query_expander.run)
g.add_node(N_SEARCHER, searcher.run)
g.add_node(N_CRITIC_Q, quality_critic.run)
g.add_node(N_SUMMARIZER, summarizer.run)
g.add_node(N_SYNTHESIZER, synthesizer.run)
g.add_node(N_CRITIC, critic.run)
g.add_node(N_TRENDS, trend_analyst.run)
# Linear edges (no branching).
g.set_entry_point(N_EXPANDER)
g.add_edge(N_EXPANDER, N_SEARCHER)
g.add_edge(N_SEARCHER, N_CRITIC_Q)
g.add_edge(N_CRITIC_Q, N_SUMMARIZER)
g.add_edge(N_SUMMARIZER, N_SYNTHESIZER)
g.add_edge(N_SYNTHESIZER, N_CRITIC)
# Conditional edge — Reflexion loop.
# After Critic: either loop back to Synthesizer or go to TrendAnalyst.
g.add_conditional_edges(
N_CRITIC,
route_after_critic,
{
N_SYNTHESIZER: N_SYNTHESIZER, # revision needed
N_TRENDS: N_TRENDS, # quality OK
},
)
g.add_edge(N_TRENDS, END)
return g.compile()
# ============================================================
# Public convenience function
# ============================================================
def run_pipeline(
topic: str,
*,
n_raw: int = 30,
top_n: int = 10,
from_year: int = DEFAULT_FROM_YEAR,
weights: dict[str, float] | None = None,
require_abstract: bool = True,
narrative_mode: bool = False,
domain_hint: str | None = DEFAULT_DOMAIN_HINT,
) -> dict[str, Any]:
"""Run the full V2 + Reflexion pipeline and return the final state.
Args:
topic: research topic to investigate.
n_raw: articles to fetch from OpenAlex.
top_n: articles to keep after QualityCritic.
from_year: earliest publication year.
weights: optional QualityScore weight overrides.
require_abstract: drop articles without abstract.
domain_hint: HF fork — biases QueryExpander's ReAct loop toward
a field (e.g. "toxicology"). Pass None to disable
and reproduce upstream (unbiased) behavior.
Returns:
Final state dict with all populated fields, including:
synthesis_iteration — how many Synthesizer runs happened
critic_feedbacks — list of CriticFeedback (one per iteration)
"""
cfg: RunConfig = {
"n_raw": n_raw,
"top_n": top_n,
"from_year": from_year,
"require_abstract": require_abstract,
}
if weights is not None:
cfg["weights"] = weights
if domain_hint is not None:
cfg["domain_hint"] = domain_hint
initial_state: dict = {
"topic": topic,
"config": cfg,
"narrative_mode": narrative_mode,
"expanded_queries": [],
"react_steps": [], # populated by QueryExpander (ReAct loop)
"synthesis_iteration": 0,
"critic_feedbacks": [],
"logs": [],
"errors": [],
}
graph = build_graph()
logger.info(
"Running pipeline on topic=%r (n_raw=%d, top_n=%d, max_reflexion=%d)",
topic, n_raw, top_n, MAX_REFLEXION_ITERATIONS,
)
final = graph.invoke(initial_state)
logger.info(
"Pipeline done. summaries=%d synthesis_iterations=%d errors=%d",
len(final.get("summaries") or []),
final.get("synthesis_iteration", 0),
len(final.get("errors") or []),
)
return final