pen-stack 7.1.1__tar.gz → 7.1.2__tar.gz

{pen_stack-7.1.1 → pen_stack-7.1.2}/CHANGELOG.md

@@ -3,6 +3,23 @@
 All notable changes to PEN-STACK are documented here. This file follows
 [Keep a Changelog](https://keepachangelog.com/).
 
+## [7.1.2] - 2026-06-25 - PEN-CHAT: chat-response latency fix
+
+### Fixed
+- Chat response time was unnecessarily high because every General-lane request paid for an Ollama embedding
+  round-trip on the query side, and the LLM call ran with a generous 150 s timeout / 450-token cap suited to the
+  engine-grounded design lane, not a textbook answer. Three minimal changes, no behaviour change for the grounded
+  lanes:
+  - `pen_stack/rag/embed.py` adds an LRU-cached `embed_query()` (256 entries) so a repeated phrasing skips the
+    embedder; `pen_stack/rag/retrieve.py` uses it.
+  - `pen_stack/web/llm_provider.py` now accepts `kind="general"`, which selects `PEN_STACK_LLM_TIMEOUT_GENERAL`
+    (45 s default, was 150 s) and `OLLAMA_NUM_PREDICT_GENERAL` / `NEMOTRON_MAX_TOKENS_GENERAL` (280 / 400, was
+    450 / 700). The default-kind path used by the engine-grounded design / explain / meta lanes is unchanged
+    (`PEN_STACK_LLM_TIMEOUT` default lowered 150 → 90).
+  - The General-lane and PEN-RAG cited-answer callers (`pen_stack/rag/ground.py`, `pen_stack/web/llm.py`) now pass
+    `kind="general"`.
+- No grounding-guard change; the LLM stays non-load-bearing and the per-lane provenance labels are unchanged.
+
 ## [7.1.1] - 2026-06-24 - PEN-CHAT: General-lane helpfulness fix + benchmark-validity correction
 
 ### Fixed

{pen_stack-7.1.1 → pen_stack-7.1.2}/CITATION.cff

@@ -1,8 +1,8 @@
 cff-version: 1.2.0
 message: "If you use PEN-STACK, please cite it as below."
 title: "PEN-STACK: open infrastructure for genome writing"
-version: 7.1.1
-date-released: 2026-06-24
+version: 7.1.2
+date-released: 2026-06-25
 authors:
   - family-names: "Mahaboob Ali"
     given-names: "Anees Ahmed"

{pen_stack-7.1.1 → pen_stack-7.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pen-stack
-Version: 7.1.1
+Version: 7.1.2
 Summary: Open infrastructure for genome writing: the Writable Genome atlas, the Writer Atlas, and the Write Planner.
 Author-email: Anees Ahmed Mahaboob Ali <ahmedaneesm@gmail.com>
 License: MIT

{pen_stack-7.1.1 → pen_stack-7.1.2}/pen_stack/__init__.py

@@ -1,2 +1,2 @@
 """PEN-STACK v3.0 - open infrastructure for genome writing."""
-__version__ = "7.1.1"
+__version__ = "7.1.2"

{pen_stack-7.1.1 → pen_stack-7.1.2}/pen_stack/rag/embed.py

@@ -10,6 +10,7 @@ from __future__ import annotations
 import json
 import os
 import urllib.request
+from functools import lru_cache
 
 import numpy as np
 
@@ -39,6 +40,24 @@ def embed_text(text: str, task: str = "query") -> np.ndarray | None:
         return None
 
 
+# v7.1.2 latency fix: cache live query embeddings in-process. The chat embeds the user's message on every General-lane
+# request; repeat phrasings ("hi", "what is DNA", "what is gene editing") would otherwise hit Ollama every time. The
+# document side is unaffected - corpus embeddings are computed once at build time and committed.
+@lru_cache(maxsize=256)
+def _embed_query_cached(text: str) -> tuple | None:
+    v = embed_text(text, task="query")
+    return tuple(float(x) for x in v) if v is not None else None
+
+
+def embed_query(text: str) -> np.ndarray | None:
+    """Cached query-side embedding. Identical to `embed_text(text, task='query')` but memoised by text."""
+    t = (text or "").strip()
+    if not t:
+        return embed_text(t, task="query")
+    cached = _embed_query_cached(t)
+    return np.asarray(cached, dtype="float32") if cached is not None else None
+
+
 def embed_corpus(texts: list[str]) -> np.ndarray:
     """Embed the corpus DOCUMENTS at BUILD time. Raises if the embedder is unreachable (the build must be
     deterministic and must never silently fall back to a different representation)."""

{pen_stack-7.1.1 → pen_stack-7.1.2}/pen_stack/rag/ground.py

@@ -90,7 +90,8 @@ def _cited_answer(message: str, r: dict, allow_llm: bool) -> dict:
         system = ("You answer ONLY from the SOURCES provided. After each claim, cite the bracketed [n] it came from. "
                   "If the sources do not cover the question, say so. Never add a number, name, or vehicle that is "
                   "not in the sources. Never present anything as a PEN-STACK-computed result.")
-        text, backend = _run_llm(f"SOURCES:\n{ctx}\n\nQUESTION: {message}\n\nAnswer concisely, citing [n].", system)
+        text, backend = _run_llm(f"SOURCES:\n{ctx}\n\nQUESTION: {message}\n\nAnswer concisely, citing [n].", system,
+                                 kind="general")  # v7.1.2 fast-path: short timeout + token cap for cited answers
         if text:
             allow = extract_grounded_numbers({"hits": [h["text"] for h in grounded_hits]})
             return {**base, "reply": _LABEL + "\n\n" + _enforce_grounding(text.strip(), allow), "backend": backend}
@@ -119,7 +120,7 @@ def ground_general(message: str, *, allow_llm: bool = True) -> dict:
         from pen_stack.web.llm import SYSTEM_GENERAL, _GENERAL_LABEL, _run_llm
         text, backend = _run_llm(
             f"USER: {msg}\n\nAnswer from general knowledge, clearly and at a graduate level. Do not present anything "
-            f"as a PEN-STACK result.", SYSTEM_GENERAL)
+            f"as a PEN-STACK result.", SYSTEM_GENERAL, kind="general")  # v7.1.2: fast-path timeout
         if text:
             return {"status": "general", "reply": _GENERAL_LABEL + "\n\n" + text.strip(), "sources": [],
                     "provenance": "general", "grounded": False, "backend": backend, "retrieval": r["method"],

{pen_stack-7.1.1 → pen_stack-7.1.2}/pen_stack/rag/retrieve.py

@@ -12,7 +12,7 @@ from functools import lru_cache
 import numpy as np
 
 from pen_stack.rag.corpus import emb_path, load_corpus
-from pen_stack.rag.embed import embed_text, lexical_scores
+from pen_stack.rag.embed import embed_query, lexical_scores
 
 
 @lru_cache(maxsize=1)
@@ -30,7 +30,7 @@ def _load():
 def retrieve(query: str, k: int = 4) -> dict:
     df, emb = _load()
     texts = df["text"].tolist()
-    qv = embed_text(query) if emb is not None else None
+    qv = embed_query(query) if emb is not None else None  # v7.1.2: LRU-cached query embedding
     if qv is not None and emb is not None:
         scores = emb @ qv  # both L2-normalised -> dot product is cosine
         method = "semantic (nomic-embed-text)"

{pen_stack-7.1.1 → pen_stack-7.1.2}/pen_stack/web/llm.py

@@ -215,10 +215,12 @@ def _angles_footer(message: str) -> str:
 # -------------------------------------------------------------------------------------- LLM backends
 # The provider backends live in pen_stack/web/llm_provider.py (PEN-CHAT P-WS2, the swappable abstraction). The chat
 # narrates through `run_llm`; the grounding guard below makes the GROUNDED result invariant to which provider fired.
-def _run_llm(prompt, system, provider=None):
-    """Delegate to the swappable provider abstraction; returns (text, backend) or (None, None)."""
+def _run_llm(prompt, system, provider=None, kind="default"):
+    """Delegate to the swappable provider abstraction; returns (text, backend) or (None, None).
+    kind="general" picks the fast-path timeout/token cap (Social / General lane); "default" keeps the longer budget
+    for engine-grounded design / explain / meta narration. v7.1.2."""
     from pen_stack.web.llm_provider import run_llm
-    return run_llm(prompt, system, provider=provider)
+    return run_llm(prompt, system, provider=provider, kind=kind)
 
 
 def _history_block(history):
@@ -316,7 +318,8 @@ def grounded_reply(message: str, history: list | None = None, *, allow_llm: bool
         # the retriever itself failed: still answer from labelled general knowledge if the LLM is on (never present
         # it as a PEN-STACK result), else point to the engine.
         if allow_llm:
-            text, backend = _run_llm(f"USER: {message}\n\nAnswer from general knowledge, clearly.", SYSTEM_GENERAL)
+            text, backend = _run_llm(f"USER: {message}\n\nAnswer from general knowledge, clearly.", SYSTEM_GENERAL,
+                                     kind="general")  # v7.1.2 fast-path: short timeout + token cap for the General lane
             if text:
                 return {"mode": "general", "provenance": "general", "grounded": False, "angles": angles,
                         "tool_results": None, "sources": [],

{pen_stack-7.1.1 → pen_stack-7.1.2}/pen_stack/web/llm_provider.py

@@ -24,18 +24,35 @@ def _ollama_base() -> str:
     return os.getenv("OLLAMA_HOST", "http://localhost:11434").rstrip("/")
 
 
-def _llm_timeout() -> float:
-    return float(os.getenv("PEN_STACK_LLM_TIMEOUT", "150"))
+# v7.1.2 latency fix: the General / Social lane never needs the long-form 450-token answer the engine-grounded lanes
+# may need, and it does not benefit from a 150 s hang waiting on a stalled provider. The kind="general" path uses a
+# shorter timeout AND a shorter num_predict; the default path (engine-grounded design / explain / meta) is unchanged.
+def _llm_timeout(kind: str = "default") -> float:
+    if kind == "general":
+        return float(os.getenv("PEN_STACK_LLM_TIMEOUT_GENERAL", "45"))
+    return float(os.getenv("PEN_STACK_LLM_TIMEOUT", "90"))
 
 
-def _call_ollama(prompt: str, system: str) -> str:
+def _ollama_num_predict(kind: str = "default") -> int:
+    if kind == "general":
+        return int(os.getenv("OLLAMA_NUM_PREDICT_GENERAL", "280"))
+    return int(os.getenv("OLLAMA_NUM_PREDICT", "450"))
+
+
+def _nemotron_max_tokens(kind: str = "default") -> int:
+    if kind == "general":
+        return int(os.getenv("NEMOTRON_MAX_TOKENS_GENERAL", "400"))
+    return int(os.getenv("NEMOTRON_MAX_TOKENS", "700"))
+
+
+def _call_ollama(prompt: str, system: str, kind: str = "default") -> str:
     import requests
     r = requests.post(
         f"{_ollama_base()}/api/generate",
         json={"model": os.getenv("OLLAMA_MODEL", "qwen2.5:3b-instruct"), "prompt": prompt, "system": system,
               "stream": False, "keep_alive": os.getenv("OLLAMA_KEEP_ALIVE", "30m"),
-              "options": {"temperature": 0.2, "num_predict": int(os.getenv("OLLAMA_NUM_PREDICT", "450"))}},
-        timeout=_llm_timeout())
+              "options": {"temperature": 0.2, "num_predict": _ollama_num_predict(kind)}},
+        timeout=_llm_timeout(kind))
     r.raise_for_status()
     return r.json()["response"]
 
@@ -48,7 +65,7 @@ def _nvidia_key() -> str | None:
     return f.read_text(encoding="utf-8").strip() if f.exists() else None
 
 
-def _call_nemotron(prompt: str, system: str) -> str:
+def _call_nemotron(prompt: str, system: str, kind: str = "default") -> str:
     import requests
     key = _nvidia_key()
     if not key:
@@ -58,7 +75,7 @@ def _call_nemotron(prompt: str, system: str) -> str:
         headers={"Authorization": f"Bearer {key}"},
         json={"model": os.getenv("NEMOTRON_MODEL", "nvidia/llama-3.3-nemotron-super-49b-v1"),
               "messages": [{"role": "system", "content": system}, {"role": "user", "content": prompt}],
-              "temperature": 0.2, "max_tokens": 700}, timeout=_llm_timeout())
+              "temperature": 0.2, "max_tokens": _nemotron_max_tokens(kind)}, timeout=_llm_timeout(kind))
     r.raise_for_status()
     return r.json()["choices"][0]["message"]["content"]
 
@@ -82,9 +99,12 @@ def default_provider() -> str:
     return _order()[0]
 
 
-def run_llm(prompt: str, system: str, provider: str | None = None) -> tuple[str | None, str | None]:
+def run_llm(prompt: str, system: str, provider: str | None = None,
+            kind: str = "default") -> tuple[str | None, str | None]:
     """Return (text, backend_name) or (None, None). If `provider` is given, use ONLY it (invariance testing);
-    otherwise try the configured order. The downstream guard makes the GROUNDED result invariant to which fired."""
+    otherwise try the configured order. `kind="general"` selects the fast-path timeout / token cap (Social / General
+    lane); `kind="default"` (the engine-grounded lanes) keeps the longer budget. The downstream guard makes the
+    GROUNDED result invariant to which provider fired."""
     if os.getenv("PEN_STACK_NO_LLM") == "1":
         return None, None
     order = [provider.strip().lower()] if provider else _order()
@@ -93,7 +113,7 @@ def run_llm(prompt: str, system: str, provider: str | None = None) -> tuple[str
         if fn is None:
             continue
         try:
-            return fn(prompt, system), name
+            return fn(prompt, system, kind=kind), name
         except Exception:  # noqa: BLE001 - a dead provider falls through to the next, else the deterministic path
             continue
     return None, None

{pen_stack-7.1.1 → pen_stack-7.1.2}/pen_stack.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pen-stack
-Version: 7.1.1
+Version: 7.1.2
 Summary: Open infrastructure for genome writing: the Writable Genome atlas, the Writer Atlas, and the Write Planner.
 Author-email: Anees Ahmed Mahaboob Ali <ahmedaneesm@gmail.com>
 License: MIT

{pen_stack-7.1.1 → pen_stack-7.1.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "pen-stack"
-version = "7.1.1"
+version = "7.1.2"
 description = "Open infrastructure for genome writing: the Writable Genome atlas, the Writer Atlas, and the Write Planner."
 readme = "README.md"
 requires-python = ">=3.11"