feat: enhance language preference handling in context building and update model retrieval in ResearchAgent

python/valuecell/agents/research_agent/core.py

@@ -6,6 +6,7 @@
 from edgar import set_identity
 from loguru import logger
 
+import valuecell.utils.model as model_utils_mod
 from valuecell.agents.research_agent.knowledge import knowledge
 from valuecell.agents.research_agent.prompts import (
     KNOWLEDGE_AGENT_EXPECTED_OUTPUT,
@@ -21,7 +22,6 @@
 from valuecell.core.agent.responses import streaming
 from valuecell.core.types import BaseAgent, StreamResponse
 from valuecell.utils.env import agent_debug_mode_enabled
-from valuecell.utils.model import get_model
 
 
 class ResearchAgent(BaseAgent):
@@ -34,7 +34,7 @@ def __init__(self, **kwargs):
             web_search,
         ]
         self.knowledge_research_agent = Agent(
-            model=get_model("RESEARCH_AGENT_MODEL_ID"),
+            model=model_utils_mod.get_model_for_agent("research_agent"),
             instructions=[KNOWLEDGE_AGENT_INSTRUCTION],
             expected_output=KNOWLEDGE_AGENT_EXPECTED_OUTPUT,
             tools=tools,

python/valuecell/agents/research_agent/prompts.py

@@ -16,6 +16,7 @@
 - fetch_event_sec_filings(ticker_or_cik, forms, start_date?, end_date?, limit?): Use this for event-driven filings like 8-K and ownership forms (3/4/5). Use date ranges and limits to control scope.
 - fetch_ashare_filings(stock_code, report_types, year?, quarter?, limit?): Use this for Chinese A-share company filings (annual reports, semi-annual reports, quarterly reports). CRITICAL: report_types parameter MUST be in English only - use "annual", "semi-annual", or "quarterly". Never use Chinese terms like "年报", "半年报", or "季报". The function will reject Chinese parameters with an error.
 - Knowledge base search: Use the agent's internal knowledge index to find summaries, historical context, analyst commentary, and previously ingested documents.
+- web_search(query): Use this for recency-sensitive information (press releases, IR pages, exchange notices), or when filings/KB lack specifics. Prefer primary sources (IR/SEC/exchanges) and reputable outlets. Encode time ranges and site filters directly in the query string (e.g., "site:investor.apple.com", "after:2025-01-01"). Always cite the exact URL(s).
 </tools>
 
 <ashare_rules>
@@ -32,6 +33,7 @@
 4. Routing by query type: see <routing_matrix> to decide filings-first vs KB-first.
 5. A-share: follow <ashare_rules> for parameter language and stock code formats.
 6. On tool failure/no results: return any partial findings you have, state the fact succinctly (e.g., "no filings returned for this window"), and propose concrete next steps (adjust window, verify ticker/CIK, increase limit).
+7. Web search (query-only): If recent events or missing context require it, use web_search with a precise query. Encode time ranges and site filters within the query itself (e.g., `site:investor.apple.com`, `after:2025-01-01`, or terms like "past 90 days"). Focus on top-quality official sources and include exact URLs in citations.
 </tool_usage_guidelines>
 
 <date_and_mapping_rules>
@@ -58,6 +60,14 @@
 3. Output style: What level of detail and technical depth is appropriate for this query?
 </response_planning>
 
+<reply_language_rules>
+- Reply in the user's language by default. If the user writes in Chinese, respond in Chinese; if in English, respond in English.
+- If the user explicitly specifies a reply language (e.g., "答复请用英文" / "please answer in Chinese"), follow that preference.
+- Do not translate API parameters or formal names that must remain in English (e.g., A-share `report_types` must be "annual", "semi-annual", "quarterly"). It's fine to explain them in the user's language.
+- Preserve numeric accuracy and unit semantics. Adapt formatting and punctuation to the user's locale when it improves readability, but do not change the underlying values.
+- Keep URLs and document titles as they are; you may add brief translated descriptors if helpful.
+</reply_language_rules>
+
 <examples>
 Example: A-share filing query (user asks "茅台2024年年报的营收是多少？"):
 Tool plan: User mentioned "年报" (annual report) in Chinese, so translate to "annual" before calling fetch_ashare_filings('600519', 'annual', year=2024).
@@ -141,6 +151,11 @@
 - Zero fabrication: if a value is unknown or not found, say so briefly and propose how to obtain it. Do not guess numbers or invent citations.
 - Strict relevance: remove tangential background; keep the response tightly scoped to the user’s ask.
 - If a blocking ambiguity exists, ask one concise clarifying question first; otherwise proceed with a reasonable default and state the assumption.
+
+Language and localization:
+- Reply in the user's language by default (e.g., Chinese input → Chinese output). Respect any explicit language preference in the prompt.
+- Keep technical parameters that must remain English as English (e.g., A-share `report_types` values), but explain their meaning in the user's language when helpful.
+- Preserve numeric fidelity; adapt units and punctuation to the locale only if unambiguous.
 </tone_and_constraints>
 
 <engagement_and_follow_up>

python/valuecell/agents/utils/context.py

@@ -31,7 +31,9 @@ def _build_lang_ctx_from_dep(
         "When composing your answer, consider the user's language and timezone:"
     )
     if user_lang:
-        parts.append(f"- Preferred language: {user_lang}")
+        parts.append(
+            f"- Preferred language: {user_lang}. Choose user's query language if different."
+        )
     else:
         parts.append(
             "- Preferred language: not set. Infer the user's language from their query and respond in that language."