Salastil/odysseus
1
0
Fork 0
mirror of https://github.com/pewdiepie-archdaemon/odysseus.git synced 2026-06-17 10:15:27 -04:00
Code Issues Packages Projects Releases Wiki Activity

fix(deep-research): wrap fetched webpage content in untrusted-context sandbox

Browse Source 
The goal-based extractor passed raw fetched webpage content straight
into the LLM prompt via string substitution, bypassing the
prompt-injection hardening layer in src/prompt_security.py.

Split EXTRACTOR_PROMPT into EXTRACTOR_SYSTEM (task instructions +
goal, trusted) and a second message built with untrusted_context_message()
(raw page content, sandboxed with <<<UNTRUSTED_SOURCE_DATA>>> guards).
This aligns the extractor with every other external-content injection
site in the codebase (agent_loop, chat_processor, chat_routes).

Fixes #3044

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Giuseppe
2026-06-06 11:37:10 +02:00
committed by GitHub
parent 893cb8254f
commit e87a1ad8d2
2 changed files with 14 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/deep_research.py
+6 -4
View File
@@ -16,7 +16,8 @@ from typing import Callable, Dict, List, Optional, Set
from src.research_utils import strip_thinking, is_low_quality from src.research_utils import strip_thinking, is_low_quality
from src.goal_based_extractor import EXTRACTOR_PROMPT from src.goal_based_extractor import EXTRACTOR_SYSTEM
from src.prompt_security import untrusted_context_message
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@@ -625,11 +626,12 @@ class DeepResearcher:
else: else:
content = truncated content = truncated
prompt = EXTRACTOR_PROMPT.format(webpage_content=content, goal=question)
try: try:
response = await self._llm( response = await self._llm(
[{"role": "user", "content": prompt}], [
{"role": "user", "content": EXTRACTOR_SYSTEM.format(goal=question)},
untrusted_context_message("webpage", content),
],
temperature=0.2, temperature=0.2,
max_tokens=2048, max_tokens=2048,
timeout=self.extraction_timeout, timeout=self.extraction_timeout,
src/goal_based_extractor.py
+8 -12
View File
@@ -3,22 +3,18 @@
Goal-based content extraction prompt inspired by Alibaba Tongyi DeepResearch. Goal-based content extraction prompt inspired by Alibaba Tongyi DeepResearch.
""" """
EXTRACTOR_PROMPT = """Please process the following webpage content and user goal to extract relevant information: EXTRACTOR_SYSTEM = """Extract relevant information from a webpage for a given research goal.
## **Webpage Content** Goal: {goal}
{webpage_content}
## **User Goal** Task guidelines:
{goal} 1. Locate the specific sections directly related to the goal within the provided webpage content.
2. Identify and extract the most relevant information; output full original context where possible, up to three or more paragraphs.
3. Organize into a concise paragraph with logical flow, judging each piece of information's contribution to the goal.
## **Task Guidelines** Respond in JSON with exactly these fields: "rational", "evidence", "summary".
1. **Content Scanning for Rational**: Locate the **specific sections/data** directly related to the user's goal within the webpage content
2. **Key Extraction for Evidence**: Identify and extract the **most relevant information** from the content, you never miss any important information, output the **full original context** of the content as far as possible, it can be more than three paragraphs.
3. **Summary Output for Summary**: Organize into a concise paragraph with logical flow, prioritizing clarity and judge the contribution of the information to the goal.
**Final Output Format using JSON format has "rational", "evidence", "summary" fields** Example:
Example output:
{{ {{
"rational": "This section discusses X which directly relates to the goal of understanding Y", "rational": "This section discusses X which directly relates to the goal of understanding Y",
"evidence": "Full quotes and context from the page...", "evidence": "Full quotes and context from the page...",