Skip to content

Personal AI, On Personal Devices

Python

docs: add design spec for Deep Research Agent v2 — SQL + LM-scan tools

Jon Saad-Falcona8f010a838a0
diff --git a/docs/superpowers/specs/2026-03-26-deep-research-agent-v2-design.md b/docs/superpowers/specs/2026-03-26-deep-research-agent-v2-design.md
new file mode 100644
index 00000000..4b238c71
--- /dev/null
+++ b/docs/superpowers/specs/2026-03-26-deep-research-agent-v2-design.md
@@ -0,0 +1,117 @@
+# Deep Research Agent v2 — Better Tools + Prompts
+
+## Goal
+
+Improve the DeepResearchAgent so it can answer aggregation queries ("who do I talk to most"), semantic/fuzzy queries ("which VCs have I spoken with"), and factual queries ("when was my trip to Spain") by adding SQL and LM-scan tools alongside better prompts.
+
+## Evaluation Criteria
+
+Re-run these 4 queries and compare to v1 (which returned empty or "no data found" for all 4):
+
+1. "When was my most recent trip to Spain?"
+2. "Which VCs have I spoken with since 2023?"
+3. "Who are the 10 people I have spoken with the most over text?"
+4. "What meetings take up most of my time based on my calendar and meeting logs?"
+
+Success = the agent produces substantive, cited answers for at least 3 of 4.
+
+## Changes
+
+### 1. New tool: `knowledge_sql`
+
+Read-only SQL queries against the `knowledge_chunks` table. The agent writes SELECT statements to aggregate, count, filter, and rank data.
+
+```python
+class KnowledgeSQLTool(BaseTool):
+ tool_id = "knowledge_sql"
+
+ def execute(self, query: str) -> ToolResult:
+ # Reject non-SELECT queries
+ # Execute against KnowledgeStore's SQLite connection
+ # Return formatted rows (max 50 rows)
+```
+
+Exposed schema (included in the tool description so the model knows the columns):
+```
+knowledge_chunks(id, content, source, doc_type, doc_id, title, author,
+ participants, timestamp, thread_id, url, metadata, chunk_index)
+```
+
+### 2. New tool: `scan_chunks`
+
+Semantic grep — pulls chunks by filter, batches through the LM with a question.
+
+```python
+class ScanChunksTool(BaseTool):
+ tool_id = "scan_chunks"
+
+ def execute(self, question: str, source: str = "", doc_type: str = "",
+ since: str = "", until: str = "", max_chunks: int = 200,
+ batch_size: int = 20) -> ToolResult:
+ # 1. Pull chunks matching filters from KnowledgeStore
+ # 2. Batch into groups of batch_size
+ # 3. For each batch, call engine.generate():
+ # "Extract info relevant to: {question}\n\nChunks:\n{batch_text}"
+ # 4. Return aggregated findings
+```
+
+This is the "LM combs through tokens" capability — catches semantic matches that BM25 keyword search misses.
+
+### 3. Wire `think` tool
+
+The `think` tool already exists at `src/openjarvis/tools/think.py`. Wire it into the DeepResearchAgent's tool list in `deep_research_setup_cmd.py`.
+
+### 4. System prompt rewrite
+
+Replace the current generic prompt with one that teaches query strategies:
+
+```
+/no_think
+You are a deep research agent with access to a personal knowledge base
+containing emails, messages, meeting notes, documents, and notes.
+
+## Your Tools
+
+- **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 SQL queries 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, aggregation, time-range queries.
+ 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 misses semantic
+ matches (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, and identify gaps.
+
+## Strategy
+
+1. Start with **think** to plan your approach — which tools suit this query?
+2. For "who/what/how many" queries → start with **knowledge_sql**
+3. For specific topics → start with **knowledge_search**
+4. If keyword search returns nothing useful → try **scan_chunks** with broader filters
+5. Cross-reference across sources — search emails, then messages, then notes
+6. After gathering evidence → write a cited narrative report
+
+## Citation Format
+
+Cite sources as: [source] title -- author
+Include a Sources section at the end.
+```
+
+### 5. Loop guard tuning
+
+Increase the `max_turns` default from 5 to 8, and configure the loop guard (if present) with a higher poll budget to accommodate the larger tool set.
+
+## What Does NOT Change
+
+- Agent class structure (DeepResearchAgent)
+- Retriever pipeline (BM25 + optional ColBERT)
+- Ingestion pipeline
+- Connectors
+- CLI command