Two-view app (chat + DB dashboard), web search, rename, boosted links

config/config.yaml

@@ -1,4 +1,4 @@
-# Awesome-WAM system configuration.
+# Awesome-Embodied&MM system configuration.
 # Everything domain/tuning-related lives here so behavior changes need no code edits.
 
 provider:
@@ -6,8 +6,8 @@ provider:
   base_url: "https://openrouter.ai/api/v1"
   api_key_env: "OPENROUTER_API_KEY"
   # Sent as OpenRouter attribution headers (optional but recommended).
-  http_referer: "https://github.com/your-org/Awesome-WAM"
-  app_title: "Awesome-WAM"
+  http_referer: "https://github.com/wzii/Awesome_Embodied_MM"
+  app_title: "Awesome-Embodied&MM"
 
 models:
   # Tiered routing. Cheap = filtering/first-pass (high volume); mid = analysis; strong =
@@ -124,6 +124,8 @@ qa:
   # passed to the qa-tier model. See stage_tiers.qa for the model.
   max_papers: 400            # cap papers in the pack (most-scored first) to bound context
   include_dropped: true      # include filtered-out papers (title-only) so the KB still covers them
+  web_search: true           # also search the web (paper original text, related concepts)
+  web_max_results: 5         # web results per query (OpenRouter web plugin; ~$0.02/query)
 
 trends:
   sim_threshold: 0.62        # cosine edge threshold for the direction graph

data/wam.db

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f7e93e5fdd9460355f30b196fe25fbf8bcf4b0f2fc6e40354e4ad6a797175324
-size 1957888
+oid sha256:afa5119bc335d2ad494cd871b265586d52be5b9803ebfd1e9c428e52b48812c7
+size 1961984

requirements.txt

@@ -8,4 +8,5 @@ requests>=2.31.0
 tenacity>=8.2.0
 networkx>=3.2
 streamlit>=1.36.0
+pandas>=2.0.0
 # No embeddings/vector DB: Q&A is long-context over a knowledge pack built from data/wam.db.

src/wam/llm/client.py

@@ -77,7 +77,7 @@ class LLMClient:
 
     # --- core call -------------------------------------------------------------
     def _call(self, model: str, messages: list[dict], *, temperature: float, max_tokens: int,
-              json_mode: bool, label: str) -> tuple[str, int, int]:
+              json_mode: bool, label: str, extra_body: dict | None = None) -> tuple[str, int, int]:
         @retry(
             reraise=True,
             stop=stop_after_attempt(self._max_retries),
@@ -90,6 +90,8 @@ class LLMClient:
                 kwargs["max_tokens"] = max_tokens
             if json_mode:
                 kwargs["response_format"] = {"type": "json_object"}
+            if extra_body:
+                kwargs["extra_body"] = extra_body
             t0 = time.monotonic()
             resp = self.client.chat.completions.create(**kwargs)
             dt = time.monotonic() - t0
@@ -109,13 +111,26 @@ class LLMClient:
     # --- public API ------------------------------------------------------------
     def complete(self, tier: str, system: str, user: str, *, temperature: float | None = None,
                  max_tokens: int | None = None, label: str = "complete",
-                 model: str | None = None) -> str:
+                 model: str | None = None, extra_body: dict | None = None) -> str:
         model = model or self.resolve_model(tier)
         temperature = self.cfg.get("models.defaults.temperature", 0.2) if temperature is None else temperature
         max_tokens = self.cfg.get("models.defaults.max_tokens", 4096) if max_tokens is None else max_tokens
         messages = [{"role": "system", "content": system}, {"role": "user", "content": user}]
         content, _, _ = self._call(model, messages, temperature=temperature,
-                                    max_tokens=max_tokens, json_mode=False, label=label)
+                                    max_tokens=max_tokens, json_mode=False, label=label,
+                                    extra_body=extra_body)
+        return content
+
+    def chat(self, tier: str, messages: list[dict], *, temperature: float | None = None,
+             max_tokens: int | None = None, label: str = "chat", model: str | None = None,
+             extra_body: dict | None = None) -> str:
+        """Multi-turn: pass a full messages list (system + prior turns + new user message)."""
+        model = model or self.resolve_model(tier)
+        temperature = self.cfg.get("models.defaults.temperature", 0.2) if temperature is None else temperature
+        max_tokens = self.cfg.get("models.defaults.max_tokens", 4096) if max_tokens is None else max_tokens
+        content, _, _ = self._call(model, messages, temperature=temperature,
+                                    max_tokens=max_tokens, json_mode=False, label=label,
+                                    extra_body=extra_body)
         return content
 
     def complete_json(self, tier: str, system: str, user: str, schema: Type[T], *,

src/wam/notify/mailer.py

@@ -227,7 +227,7 @@ def build_html(cfg: Config, conn: sqlite3.Connection, today: str | None = None)
 
     S = "font-family:-apple-system,Segoe UI,Roboto,Helvetica,Arial,sans-serif"
     parts = [f'<div style="max-width:640px;margin:auto;{S};color:#1a1a1a">']
-    parts.append(f'<h1 style="font-size:20px;margin:0 0 8px">🤖 WAM Daily — {today}</h1>')
+    parts.append(f'<h1 style="font-size:20px;margin:0 0 8px">🤖 Embodied&amp;MM Daily — {today}</h1>')
     new_adj = conn.execute("SELECT count(*) FROM papers WHERE track='adjacent' AND first_seen=?",
                            (today,)).fetchone()[0]
     total_core, total_adj = counts.get("core", 0), counts.get("adjacent", 0)
@@ -288,9 +288,9 @@ def build_html(cfg: Config, conn: sqlite3.Connection, today: str | None = None)
         'WAM (bold = weighted 2×): spd=inference speed, gen=generalist, spec=specialist, '
         'cost=inference cost, trust=trustworthiness, collab=collaborative, '
         'ctrl=controlled generation. “–” = the paper doesn’t address it.</p>')
-    parts.append(f'<p style="color:#999;font-size:12px;margin-top:8px">Awesome-WAM · '
-                 f'<a href="https://github.com/your-org/Awesome-WAM">repo</a></p></div>')
-    subject = f"WAM Daily — {today}: {new_core} new" + ("" if not fallback else " (recap)")
+    parts.append(f'<p style="color:#999;font-size:12px;margin-top:8px">Awesome-Embodied&amp;MM · '
+                 f'<a href="https://github.com/wzii/Awesome_Embodied_MM">repo</a></p></div>')
+    subject = f"Embodied&MM Daily — {today}: {new_core} new" + ("" if not fallback else " (recap)")
     return subject, "".join(parts)
 
 

src/wam/pipeline/links.py

@@ -14,7 +14,7 @@ import sqlite3
 
 from wam.config import Config
 from wam.logging import get_logger
-from wam.sources import pdf
+from wam.sources import papers_with_code, pdf
 
 log = get_logger("pipeline.links")
 
@@ -31,6 +31,11 @@ def _clean(url: str) -> str:
     return url.rstrip(").,;:'\"]}").rstrip("/")
 
 
+def _prep(text: str) -> str:
+    """Rejoin URLs broken across PDF lines: drop hyphenated breaks, unwrap newlines."""
+    return text.replace("-\n", "").replace("\n", " ")
+
+
 def _best(text: str, pattern: re.Pattern, deny: tuple = ()) -> str | None:
     cands = [(m.start(), _clean(m.group(0))) for m in pattern.finditer(text)]
     cands = [(p, u) for p, u in cands if not any(d in u.lower() for d in deny)]
@@ -44,12 +49,12 @@ def _best(text: str, pattern: re.Pattern, deny: tuple = ()) -> str | None:
 
 
 def run(cfg: Config, conn: sqlite3.Connection, limit: int | None = None,
-        download: bool = True) -> int:
+        download: bool = True, use_pwc: bool = True) -> int:
     q = ("SELECT id, abstract, links_json FROM papers WHERE track IN ('core','adjacent')")
     if limit:
         q += f" LIMIT {int(limit)}"
     rows = conn.execute(q).fetchall()
-    log.info("links: scanning %d papers (download=%s)", len(rows), download)
+    log.info("links: scanning %d papers (download=%s, pwc=%s)", len(rows), download, use_pwc)
     found = 0
     for i, r in enumerate(rows):
         links = json.loads(r["links_json"] or "{}")
@@ -57,21 +62,24 @@ def run(cfg: Config, conn: sqlite3.Connection, limit: int | None = None,
             continue
         pdf_url = links.get("pdf") or (f"https://arxiv.org/pdf/{r['id'].split(':',1)[1]}"
                                        if r["id"].startswith("arxiv:") else "")
-        text = pdf.get_text(cfg, r["id"], pdf_url if download else "", max_chars=60000)
-        blob = (text or "") + "\n" + (r["abstract"] or "")
-        if not blob.strip():
-            continue
+        text = pdf.get_text(cfg, r["id"], pdf_url if download else "", max_chars=80000)
+        blob = _prep((text or "") + "\n" + (r["abstract"] or ""))  # rejoin line-broken URLs
         changed = False
-        if not links.get("code"):
-            gh = _best(blob, _GH, _GH_DENY)
-            if gh:
-                links["code"] = gh
-                changed = True
-        if not links.get("hf"):
-            hf = _best(blob, _HF)
-            if hf:
-                links["hf"] = hf
-                changed = True
+        if blob.strip():
+            if not links.get("code"):
+                gh = _best(blob, _GH, _GH_DENY)
+                if gh:
+                    links["code"], changed = gh, True
+            if not links.get("hf"):
+                hf = _best(blob, _HF)
+                if hf:
+                    links["hf"], changed = hf, True
+        # Fallback: Papers-with-Code (canonical arXiv->repo) when still no code link.
+        if use_pwc and not links.get("code") and r["id"].startswith("arxiv:"):
+            url = papers_with_code._repo_for_arxiv(r["id"].split(":", 1)[1],
+                                                   cfg.get("constants.request_timeout", 90))
+            if url:
+                links["code"], changed = _clean(url), True
         if changed:
             conn.execute("UPDATE papers SET links_json=? WHERE id=?",
                          (json.dumps(links), r["id"]))

src/wam/render/readme.py

@@ -180,7 +180,7 @@ def link_integrity(conn: sqlite3.Connection) -> list[str]:
 def render_readme(cfg: Config, conn: sqlite3.Connection, today: str | None = None) -> str:
     today = today or date.today().isoformat()
     c = _counts(conn)
-    md = f"""# Awesome-WAM
+    md = f"""# Awesome-Embodied&MM
 
 > Daily-updated intelligence on **World Action Models** — world models, vision-language-action
 > (VLA) models, action-conditioned video/world generation, robot foundation models, and
@@ -210,7 +210,7 @@ _Not scored; surfaced for techniques transferable to WAM._
 ## 📰 Embodied / Physical-AI News
 {_news(conn)}
 ---
-_Generated by [Awesome-WAM](https://github.com/your-org/Awesome-WAM)._
+_Generated by [Awesome-Embodied&MM](https://github.com/wzii/Awesome_Embodied_MM)._
 """
     (cfg.root / "README.md").write_text(md, encoding="utf-8")
     log.info("wrote README.md")
@@ -227,7 +227,7 @@ def render_digest(cfg: Config, conn: sqlite3.Connection, today: str | None = Non
         "SELECT title, links_json, scores_json, summary_json FROM papers WHERE track='core' "
         "AND first_seen=? AND scores_json IS NOT NULL "
         "ORDER BY json_extract(scores_json,'$.weighted_total') DESC LIMIT 15", (today,)).fetchall()
-    lines = [f"# WAM Daily Digest — {today}\n",
+    lines = [f"# Embodied&MM Daily Digest — {today}\n",
              f"**New today:** {new_core} core · {new_adj} adjacent papers\n",
              "## Top new papers\n"]
     if not rows:

src/wam/sources/pdf.py

@@ -39,7 +39,7 @@ def get_text(cfg: Config, paper_id: str, pdf_url: str, *, max_chars: int = 60000
         return None
     try:
         resp = requests.get(pdf_url, timeout=cfg.get("constants.request_timeout", 90),
-                            headers={"User-Agent": "Awesome-WAM/0.1"})
+                            headers={"User-Agent": "Awesome-Embodied-MM/0.1"})
         resp.raise_for_status()
         with fitz.open(stream=resp.content, filetype="pdf") as doc:
             text = "\n".join(page.get_text() for page in doc)

src/wam/store/benchmarks.py

@@ -46,9 +46,11 @@ def insert_benchmark(conn: sqlite3.Connection, row: BenchmarkRow, source_paper_i
             metric_value, inference_speed, speed_unit, inference_cost, cost_unit, hardware,
             source_paper_id, claimed_by_authors, notes, extracted_on)
            VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)""",
-        (vk, row.model_name, row.training_dataset, row.benchmark, row.task, row.split,
-         row.metric_name, row.metric_value, row.inference_speed, row.speed_unit,
-         row.inference_cost, row.cost_unit, row.hardware, source_paper_id,
+        # Coalesce the UNIQUE-key text cols to '' so SQLite dedups them (NULLs are all
+        # distinct in a UNIQUE index, which would let near-duplicate rows slip through).
+        (vk, row.model_name, row.training_dataset, row.benchmark, row.task or "",
+         row.split or "", row.metric_name or "", row.metric_value, row.inference_speed,
+         row.speed_unit, row.inference_cost, row.cost_unit, row.hardware, source_paper_id,
          int(row.claimed_by_authors), row.notes, date.today().isoformat()),
     )
 

src/wam/webapp/app.py

@@ -1,10 +1,10 @@
-"""Streamlit Q&A app — the WAM knowledge base front-end (long-context, no RAG).
+"""Streamlit app — Awesome-Embodied&MM: a chat assistant + a database dashboard.
+
+Two views (sidebar): 💬 Chat (multi-turn, long-context Q&A grounded in the corpus + live web
+search) and 📊 Database (visual browse of papers, scores, benchmarks, directions, authors).
 
 Run locally:   streamlit run src/wam/webapp/app.py
 On HF Spaces:  set OPENROUTER_API_KEY as a Space secret; this file is the entry point.
-
-The OpenRouter key stays server-side. Answers are grounded in a knowledge pack built from the
-committed data/wam.db (paper summaries + scores + leaderboard + authors + trends).
 """
 
 from __future__ import annotations
@@ -17,26 +17,54 @@ sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
 import streamlit as st  # noqa: E402
 
 from wam.config import load_config  # noqa: E402
-from wam.webapp.qa import answer  # noqa: E402
-
-st.set_page_config(page_title="Awesome-WAM Knowledge Base", page_icon="🤖", layout="centered")
-st.title("🤖 Awesome-WAM Knowledge Base")
-st.caption("Ask about World Action Models — papers, benchmarks, authors, trends. "
-           "Answers are grounded in the tracked corpus and cite paper ids.")
 
+st.set_page_config(page_title="Awesome-Embodied&MM", page_icon="🤖", layout="wide")
 cfg = load_config()
 db_exists = cfg.path("db").exists()
+
+st.sidebar.title("🤖 Awesome-Embodied&MM")
+view = st.sidebar.radio("View", ["💬 Chat", "📊 Database"], label_visibility="collapsed")
 if not db_exists:
-    st.warning("No data found yet. Run the pipeline to populate data/wam.db.")
-
-examples = ["Which VLA models report real-time inference?",
-            "What are the rising research directions?",
-            "Who works on world models for autonomous driving?"]
-q = st.text_input("Your question", placeholder=examples[0])
-st.caption("Try: " + " · ".join(f"_{e}_" for e in examples))
-
-if q and db_exists:
-    with st.spinner("Reading the corpus and composing an answer…"):
-        res = answer(q, cfg=cfg)
-    st.markdown(res["answer"])
-    st.caption(f"Grounded in a {res['pack_chars']:,}-char knowledge pack.")
+    st.sidebar.warning("No data yet — run the pipeline to populate data/wam.db.")
+
+
+def _chat() -> None:
+    from wam.webapp.qa import answer
+    st.title("💬 Chat")
+    st.caption("Ask about World Action Models, VLA, world models & video generation. Grounded "
+               "in the tracked corpus (cited by paper id) + live web search for original text "
+               "and related concepts.")
+    with st.sidebar:
+        if st.button("🗑 Clear chat"):
+            st.session_state.messages = []
+            st.rerun()
+    st.session_state.setdefault("messages", [])
+    for m in st.session_state.messages:
+        with st.chat_message(m["role"]):
+            st.markdown(m["content"])
+    if q := st.chat_input("Ask a question…", disabled=not db_exists):
+        st.session_state.messages.append({"role": "user", "content": q})
+        with st.chat_message("user"):
+            st.markdown(q)
+        with st.chat_message("assistant"):
+            with st.spinner("Thinking (corpus + web)…"):
+                res = answer(q, history=st.session_state.messages[:-1], cfg=cfg)
+            st.markdown(res["answer"])
+            web = " + live web" if res.get("web") else ""
+            st.caption(f"Grounded in a {res.get('pack_chars', 0):,}-char pack{web}.")
+        st.session_state.messages.append({"role": "assistant", "content": res["answer"]})
+
+
+def _database() -> None:
+    from wam.webapp import dashboard
+    st.title("📊 Database")
+    if db_exists:
+        dashboard.render(cfg)
+    else:
+        st.info("No data yet.")
+
+
+if view.startswith("💬"):
+    _chat()
+else:
+    _database()

src/wam/webapp/dashboard.py

@@ -0,0 +1,108 @@
+"""Database dashboard — visual browse of the Awesome-Embodied&MM corpus (Streamlit).
+
+Read-only views over data/wam.db: headline metrics, research directions (chart), top scored
+papers with their two-layer dimensions, the benchmark leaderboard, and influential authors.
+"""
+
+from __future__ import annotations
+
+import json
+import sqlite3
+
+import pandas as pd
+import streamlit as st
+
+from wam.config import Config
+
+_WAM_TOP4 = [("inference_speed", "spd"), ("generalist", "gen"),
+             ("specialist", "spec"), ("inference_cost", "cost")]
+
+
+def _conn(cfg: Config) -> sqlite3.Connection:
+    return sqlite3.connect(f"file:{cfg.path('db')}?mode=ro", uri=True)
+
+
+def render(cfg: Config) -> None:
+    con = _conn(cfg)
+
+    counts = dict(con.execute("SELECT track, count(*) FROM papers GROUP BY track").fetchall())
+    nbench = con.execute("SELECT count(*) FROM benchmarks").fetchone()[0]
+    nvar = con.execute("SELECT count(*) FROM model_variants").fetchone()[0]
+    nauth = con.execute("SELECT count(*) FROM authors").fetchone()[0]
+    c = st.columns(6)
+    c[0].metric("Core", counts.get("core", 0))
+    c[1].metric("Adjacent", counts.get("adjacent", 0))
+    c[2].metric("News", counts.get("news", 0))
+    c[3].metric("Benchmark rows", nbench)
+    c[4].metric("Model variants", nvar)
+    c[5].metric("Authors", nauth)
+
+    # --- Research directions ---
+    st.subheader("📈 Research directions")
+    fronts = pd.read_sql_query(
+        "SELECT name, size, momentum, summary FROM fronts WHERE snapshot_date="
+        "(SELECT max(snapshot_date) FROM fronts) ORDER BY size DESC", con)
+    if not fronts.empty:
+        st.bar_chart(fronts.set_index("name")["size"], height=320)
+        st.dataframe(fronts, use_container_width=True, hide_index=True)
+    else:
+        st.info("No trend snapshot yet.")
+
+    # --- Top papers ---
+    st.subheader("🏆 Top scored papers")
+    track = st.selectbox("Track", ["core", "adjacent", "all"], index=0)
+    where = "" if track == "all" else f"AND track='{track}'"
+    rows = con.execute(
+        f"SELECT id, title, track, published, relevance, scores_json, links_json FROM papers "
+        f"WHERE scores_json IS NOT NULL {where} "
+        f"ORDER BY json_extract(scores_json,'$.weighted_total') DESC LIMIT 200").fetchall()
+    recs = []
+    for pid, title, trk, pub, rel, sj, lj in rows:
+        s = json.loads(sj or "{}")
+        wam = s.get("wam", {})
+        links = json.loads(lj or "{}")
+        rec = {"score": s.get("weighted_total"), "title": title, "track": trk,
+               "published": pub}
+        for key, lbl in _WAM_TOP4:
+            v = wam.get(key)
+            rec[lbl] = v if isinstance(v, int) else None
+        rec["code"] = links.get("code") or ""
+        rec["abs"] = links.get("abs") or ""
+        recs.append(rec)
+    if recs:
+        df = pd.DataFrame(recs)
+        st.dataframe(df, use_container_width=True, hide_index=True, column_config={
+            "abs": st.column_config.LinkColumn("abs"),
+            "code": st.column_config.LinkColumn("code")})
+        st.caption("Score = weighted total. spd/gen/spec/cost = top-4 WAM dims (blank = N/A).")
+    else:
+        st.info("No scored papers for this track yet.")
+
+    # --- Benchmark leaderboard ---
+    st.subheader("📊 Benchmark leaderboard")
+    bench = pd.read_sql_query(
+        "SELECT benchmark, task, model_name AS model, training_dataset AS trained_on, "
+        "metric_name AS metric, metric_value AS value, "
+        "CASE claimed_by_authors WHEN 1 THEN 'authors' ELSE '3rd-party' END AS source "
+        "FROM benchmarks WHERE metric_value IS NOT NULL ORDER BY benchmark, value DESC", con)
+    if not bench.empty:
+        st.caption("Model identity = (model, training data) — same name on different data is a "
+                   "distinct row.")
+        st.dataframe(bench, use_container_width=True, hide_index=True)
+    else:
+        st.info("No benchmark rows yet.")
+
+    # --- Authors ---
+    st.subheader("👥 Influential authors")
+    arows = con.execute(
+        "SELECT name, citations, h_index, paper_ids_json, directions, s2_url FROM authors "
+        "ORDER BY json_array_length(paper_ids_json) DESC").fetchall()
+    if arows:
+        adf = pd.DataFrame([{
+            "author": n, "papers": len(json.loads(pj or "[]")), "citations": cit,
+            "h_index": h, "directions": (d or "")[:200], "s2": url or ""}
+            for n, cit, h, pj, d, url in arows])
+        st.dataframe(adf, use_container_width=True, hide_index=True,
+                     column_config={"s2": st.column_config.LinkColumn("s2")})
+    else:
+        st.info("No authors yet.")

src/wam/webapp/qa.py

@@ -19,10 +19,14 @@ from wam.store import Database
 log = get_logger("webapp.qa")
 
 SYSTEM = (
-    "You answer questions about World Action Models research using ONLY the knowledge pack "
-    "below (tracked papers with scores, a benchmark leaderboard, author directions, and "
-    "trends). Cite papers by their id like (arxiv:2606.01234). If the pack doesn't contain "
-    "the answer, say so — never invent papers, numbers, or citations. Be concise and concrete.")
+    "You answer questions about World Action Models research. You have two sources:\n"
+    "1) the KNOWLEDGE PACK below — our curated corpus (tracked papers with scores, a "
+    "benchmark leaderboard, author directions, trends). Use it for facts about what we track.\n"
+    "2) WEB SEARCH results (when provided) — use them for the original paper text, definitions "
+    "of related concepts, and anything not in the pack.\n"
+    "Cite tracked papers by id like (arxiv:2606.01234) and web sources by their URL. Prefer the "
+    "pack for our corpus facts; never invent papers, numbers, or citations. Be concise and "
+    "concrete, and write in English.")
 
 
 def build_pack(conn: sqlite3.Connection, max_papers: int = 400,
@@ -88,7 +92,12 @@ def build_pack(conn: sqlite3.Connection, max_papers: int = 400,
     return "\n\n".join(sections)
 
 
-def answer(question: str, cfg: Config | None = None, client: LLMClient | None = None) -> dict:
+def answer(question: str, history: list[dict] | None = None, cfg: Config | None = None,
+           client: LLMClient | None = None) -> dict:
+    """Answer a question, optionally with prior conversation turns for multi-turn chat.
+
+    ``history`` is a list of {"role": "user"|"assistant", "content": ...} from earlier turns.
+    """
     cfg = cfg or load_config()
     client = client or LLMClient(cfg)
     with Database(cfg) as db:
@@ -96,8 +105,15 @@ def answer(question: str, cfg: Config | None = None, client: LLMClient | None =
                           include_dropped=bool(cfg.get("qa.include_dropped", True)))
     if not pack.strip():
         return {"answer": "The knowledge base is empty — run the pipeline first.", "pack_chars": 0}
-    # Generous ceiling: the qa-tier model may be a reasoning model (burns tokens before the
-    # answer); too low a cap yields empty content.
-    ans = client.complete("qa", SYSTEM, f"KNOWLEDGE PACK:\n{pack}\n\nQUESTION: {question}",
-                          label="qa", max_tokens=6000)
-    return {"answer": ans, "pack_chars": len(pack)}
+    # Optionally let the model also search the web (paper original text, related concepts)
+    # via OpenRouter's web plugin — works with any model.
+    extra_body = None
+    if bool(cfg.get("qa.web_search", True)):
+        extra_body = {"plugins": [{"id": "web",
+                                   "max_results": int(cfg.get("qa.web_max_results", 5))}]}
+    messages = [{"role": "system", "content": f"{SYSTEM}\n\nKNOWLEDGE PACK:\n{pack}"}]
+    messages += list(history or [])
+    messages.append({"role": "user", "content": question})
+    # Generous ceiling: the qa-tier model may be a reasoning model (burns tokens first).
+    ans = client.chat("qa", messages, label="qa", max_tokens=6000, extra_body=extra_body)
+    return {"answer": ans, "pack_chars": len(pack), "web": bool(extra_body)}