clawverse/clawlibrary
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Make summarize fast: abstract-based input (~7s/paper vs ~2min)

Browse Source 
Feeding 16k chars of body to an 8B model took 126s/paper (42h for the vault).
The abstract already carries the gist, so cap the body excerpt to 2.5k chars
and num_predict to 500 -> ~7s/paper (~2.3h total), still valid structured JSON.
Add --model flag (default gemma4:E4B).

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
This commit is contained in:
Omar Sobh
2026-06-03 19:37:53 -05:00
co-authored by Claude Opus 4.8
parent c1da5ad249
commit 0706d18799
2 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
clawlibrary/__main__.py
+2 -1
View File
@@ -183,7 +183,7 @@ def cmd_summarize(args: argparse.Namespace) -> int:
from .summarize import append_insight, has_summary, summarize_one from .summarize import append_insight, has_summary, summarize_one
config, _ = _load(args.config, args.topics) config, _ = _load(args.config, args.topics)
oc = OllamaClient() oc = OllamaClient(text_model=args.model)
if not oc.is_up(): if not oc.is_up():
console.print("[red]Ollama not reachable.[/red]") console.print("[red]Ollama not reachable.[/red]")
return 2 return 2
@@ -313,6 +313,7 @@ def build_parser() -> argparse.ArgumentParser:
pm = sub.add_parser("summarize", help="LLM structured summaries (gemma4 via Ollama)") pm = sub.add_parser("summarize", help="LLM structured summaries (gemma4 via Ollama)")
pm.add_argument("--limit", type=int, default=None) pm.add_argument("--limit", type=int, default=None)
pm.add_argument("--force", action="store_true", help="re-summarize even if done") pm.add_argument("--force", action="store_true", help="re-summarize even if done")
pm.add_argument("--model", default="gemma4:E4B", help="Ollama model for summaries")
pm.set_defaults(func=cmd_summarize) pm.set_defaults(func=cmd_summarize)
pb = sub.add_parser("embed", help="embed papers for semantic search (nomic via Ollama)") pb = sub.add_parser("embed", help="embed papers for semantic search (nomic via Ollama)")
clawlibrary/summarize.py
+4 -2
View File
@@ -19,7 +19,9 @@ PROJECT_CONTEXT = (
"and quantum computing/hardware/networking." "and quantum computing/hardware/networking."
) )
MAX_TEXT_CHARS = 16000 # The abstract carries the gist; a short body excerpt grounds problem/method.
# Keeping the input small is what makes summarization ~7s/paper instead of ~2min.
MAX_TEXT_CHARS = 2500
_PROMPT = """You are a research assistant building an insight digest. _PROMPT = """You are a research assistant building an insight digest.
@@ -51,7 +53,7 @@ def build_prompt(title: str, abstract: str, body: str, context: str = PROJECT_CO
def summarize_paper(client: OllamaClient, *, title: str, abstract: str, body: str) -> dict: def summarize_paper(client: OllamaClient, *, title: str, abstract: str, body: str) -> dict:
"""Return a structured summary dict (best-effort parse of the model JSON).""" """Return a structured summary dict (best-effort parse of the model JSON)."""
raw = client.generate(build_prompt(title, abstract, body), json_mode=True, num_predict=800) raw = client.generate(build_prompt(title, abstract, body), json_mode=True, num_predict=500)
try: try:
data = json.loads(raw) data = json.loads(raw)
except (json.JSONDecodeError, TypeError): except (json.JSONDecodeError, TypeError):