Personal AI, On Personal Devices
Python
feat: rewrite DeepResearch system prompt, wire SQL + scan + think tools, increase max_turns to 8
Jon Saad-Falcona1626960c1dd
diff --git a/src/openjarvis/agents/deep_research.py b/src/openjarvis/agents/deep_research.pyindex 64792319..0d7c4a32 100644--- a/src/openjarvis/agents/deep_research.py+++ b/src/openjarvis/agents/deep_research.py@@ -32,25 +32,48 @@ def _tc_args(tc: dict) -> str:DEEP_RESEARCH_SYSTEM_PROMPT = """\/no_think-You are a deep research agent. Your job is to search the user's personal \-knowledge base thoroughly and produce a well-cited narrative report.+You are a deep research agent with access to a personal knowledge base \+containing emails, messages, meeting notes, documents, and notes.-## Instructions+## Your Tools-1. Use the **knowledge_search** tool to search ingested personal data \-(emails, Slack messages, documents, notes). Make multiple searches with \-different queries to cross-reference information across sources.-2. Use the **think** tool to reason between searches -- plan your next \-query, evaluate what you have found so far, and identify gaps.-3. After gathering enough evidence, produce a final narrative answer \-with inline citations in the format: [source] title -- author-4. Include a **Sources** section at the end listing all referenced sources.+- **knowledge_search**: BM25 keyword search. Best for finding specific topics, \+names, or phrases. Use filters (source, author, since, until) to narrow results.++- **knowledge_sql**: Run read-only SQL against the knowledge_chunks table. \+Schema: knowledge_chunks(id, content, source, doc_type, doc_id, title, \+author, participants, timestamp, thread_id, url, metadata, chunk_index). \+Best for counting, ranking, and aggregation. \+Example: SELECT author, COUNT(*) as n FROM knowledge_chunks \+WHERE source='imessage' GROUP BY author ORDER BY n DESC LIMIT 10++- **scan_chunks**: Semantic search — feeds chunks to an LM that reads the \+actual text looking for relevant information. Use when keyword search returns \+nothing useful or when you need semantic matching (e.g. searching for 'VCs' \+when text says 'fundraising round'). Slower but catches what BM25 misses.++- **think**: Reasoning scratchpad. Use between searches to plan your next \+query, evaluate findings so far, and identify gaps.++## Strategy++1. Start with **think** to plan your approach — which tools suit this query?+2. For "who/what/how many" queries → use **knowledge_sql** with GROUP BY+3. For specific topics or names → use **knowledge_search**+4. If keyword search returns nothing useful → try **scan_chunks** with filters+5. Cross-reference across sources (emails, messages, meeting notes)+6. After gathering evidence → write a cited narrative report++## Citation Format++Cite sources as: [source] title -- author+End with a Sources section listing all referenced items.## Rules- Always cite your sources. Never present information without attribution.- Make at least two searches to cross-reference across different sources.-- If a search returns no results, try rephrasing the query.+- If a search returns no results, try a different tool or rephrase the query.- Prefer specificity: filter by source, author, or date when appropriate.- Your final answer should be a coherent narrative, not a list of raw results."""@@ -60,7 +83,7 @@ class DeepResearchAgent(ToolUsingAgent):"""Multi-hop research agent with native function calling and citations."""agent_id = "deep_research"- _default_max_turns = 5+ _default_max_turns = 8_default_temperature = 0.3_default_max_tokens = 4096diff --git a/src/openjarvis/cli/deep_research_setup_cmd.py b/src/openjarvis/cli/deep_research_setup_cmd.pyindex 99964f11..5e9af601 100644--- a/src/openjarvis/cli/deep_research_setup_cmd.py+++ b/src/openjarvis/cli/deep_research_setup_cmd.py@@ -267,6 +267,9 @@ def _launch_chat(store: KnowledgeStore, console: Console) -> None:from openjarvis.connectors.retriever import TwoStageRetrieverfrom openjarvis.engine.ollama import OllamaEnginefrom openjarvis.tools.knowledge_search import KnowledgeSearchTool+ from openjarvis.tools.knowledge_sql import KnowledgeSQLTool+ from openjarvis.tools.scan_chunks import ScanChunksTool+ from openjarvis.tools.think import ThinkToolconsole.print("\n[bold]Setting up Deep Research agent...[/bold]")@@ -290,20 +293,26 @@ def _launch_chat(store: KnowledgeStore, console: Console) -> None:)return- # Retriever + tool+ # Toolsretriever = TwoStageRetriever(store)- search_tool = KnowledgeSearchTool(retriever=retriever)+ tools = [+ KnowledgeSearchTool(retriever=retriever),+ KnowledgeSQLTool(store=store),+ ScanChunksTool(store=store, engine=engine, model=_OLLAMA_MODEL),+ ThinkTool(),+ ]# Agentagent = DeepResearchAgent(engine=engine,model=_OLLAMA_MODEL,- tools=[search_tool],+ tools=tools,interactive=True,)console.print(f"[green]Ready![/green] Using [bold]{_OLLAMA_MODEL}[/bold] via Ollama.\n"+ "Tools: knowledge_search, knowledge_sql, scan_chunks, think\n""Type your research question. Type [bold]/quit[/bold] to exit.\n")diff --git a/tests/agents/test_deep_research.py b/tests/agents/test_deep_research.pyindex 31b0400b..3c712872 100644--- a/tests/agents/test_deep_research.py+++ b/tests/agents/test_deep_research.py@@ -173,6 +173,6 @@ def test_agent_system_prompt_mentions_research(mock_engine, store):def test_agent_defaults():"""Verify agent_id, default max_turns, temperature, max_tokens."""assert DeepResearchAgent.agent_id == "deep_research"- assert DeepResearchAgent._default_max_turns == 5+ assert DeepResearchAgent._default_max_turns == 8assert DeepResearchAgent._default_temperature == 0.3assert DeepResearchAgent._default_max_tokens == 4096