This page walks through the live smoke scripts and the full agent. Every example uses the hello_<cap>_demo profile so it can't clobber real data.

0. The one-liner: full agent smoke

DATAMIND__LLM__API_BASE=http://35.220.164.252:3888 \
DATAMIND__LLM__API_KEY=sk-... \
python -m datamind.scripts.hello_agent

It seeds a profile with a KB, graph, and SQLite, then asks four real questions in Chinese:

Q	Tools the agent picked (autonomously)	Result
Status meeting 什么时候开？	memory_recall → kb_search	"周一 14:00 Shanghai time" from company_handbook.md
Search platform 负责人是谁？他在哪个城市？	kb_search → graph_search_entities → graph_neighbors ×2	"Ann leads Search platform; city is unknown from the graph" (honest — city is in SQLite, not the graph)
工程部 Shanghai 员工工资加起来是多少？	db_query_nl → db_list_tables → db_describe_table → db_query_sql ×2	¥26,000 with a formatted table (recovers after a wrong first SQL)
帮我记住下周三会议调到周四	memory_save	Persisted with metadata

Highlights:

• Autonomous tool selection — no hints, no manual routing.
• Graceful error recovery — if NL2SQL generates the wrong column, the agent reads the schema and retries.
• Chinese output, English tool names — the split we designed.

1. KB (RAG) — hello_kb

python -m datamind.scripts.hello_kb

Seeds a two-file corpus (ai_history.md, rag.md), indexes it with real embeddings, then runs three queries. Example output:

[hello_kb] indexed: {'pre_chunked': 0, 'raw_chunks': 4, 'total_embedded': 4}

[hello_kb] Q: When did AI research start?
  - score=0.033 source='ai_history.md'  'Artificial intelligence research began in the 1950s…'

[hello_kb] Q: What is reciprocal rank fusion?
  - score=0.033 source='rag.md'  'Hybrid retrieval mixes dense vector search with BM25…'

Strategies are configurable via DATAMIND__RETRIEVAL__STRATEGY:

Strategy	When to use
simple	Cheapest; good baseline
multi_query	Ambiguous queries — LLM rewrites into N siblings
hybrid	Default; BM25 + vector fused with RRF (best on most corpora)

2. Graph — hello_graph

python -m datamind.scripts.hello_graph

No network. Seeds 8 triples about Ann / Bob / Acme / Shanghai, then demonstrates:

• Entity search (case-insensitive + fuzzy)
• 3-hop traversal: Ann → Acme → Shanghai → China
• Relation filter: pass ["works_at", "located_in", "in_country"] to restrict edges followed
• Direction-aware neighbors
• Persistence: the same graph reloads from storage/hello_graph_demo/graph.json on next run

3. Database — hello_db

python -m datamind.scripts.hello_db

Creates a SQLite demo (employees + projects), exercises every DB tool:

• db_list_tables / db_describe_table
• db_query_sql: group-by + aggregates
• Safeguard: DELETE FROM employees is rejected with DestructiveSQLError
• db_query_nl: real NL2SQL via the gateway — generates JOINs and filters

Example:

[hello_db] db_query_sql (group by department):
  rows=[['Eng', 3, 12666.67], ['Sales', 1, 9000.0], ['HR', 1, 8500.0]]

[hello_db] destructive rejected OK: …

[hello_db] db_query_nl:
  generated SQL: SELECT name, salary FROM employees WHERE city='Shanghai' AND department='Eng'
  rows: [['Ann', 15000], ['Bob', 11000], ['Evan', 15000]]

4. Skills — hello_skills

python -m datamind.scripts.hello_skills

Discovers every .claude/skills/<name>/SKILL.md manifest, indexes them against live embeddings, then searches semantically:

skill_search 'how should I review a pull request?'
  score=0.513  name=code-review
  score=0.194  name=db-ops-sop

skill_search '慢查询怎么排查'
  score=0.392  name=db-ops-sop
  score=0.330  name=code-review

Also exercises the code skills: calculator("2 * (3 + sqrt(16))") → 14, 100°C → 212°F.

5. Memory — hello_memory

python -m datamind.scripts.hello_memory

Shows all three layers:

• Short-term rolling window (in-memory FIFO)
• Long-term SQLite + cosine recall
• LLM fact extraction:

[hello_memory] extracted 2 fact(s):
  - User is moving to Shenzhen next month
  - User plans to start jogging in the morning

from the raw turn "I'm moving to Shenzhen next month and will start jogging in the morning."

6. Enterprise demo — hello_enterprise (recommended)

The flagship v0.2 demo: a medium-sized realistic profile + 8 cross-backend complex questions.

6.1 Seed once

python -m datamind.scripts.seed_enterprise_demo

Sets up:

• 17 KB documents (employee handbook, security policy, incident SOP, 3 product architectures, API ref, quarterly retros, OKRs, culture…)
• 64 graph nodes / 91 edges (org chart / project deps / product components / incident → service)
• 6 SQL tables / 101 rows (departments / employees / projects / project_members / incidents / performance_reviews)

6.2 Run 8 cross-backend questions

DATAMIND__DATA__PROFILE=enterprise_demo \
  python -m datamind.scripts.hello_enterprise

Coverage:

Question	Capabilities exercised
What's the deploy window? what's the pre-deploy checklist?	KB multi-doc
2025 Q4 incidents on Search Platform? responder H2 perf scores?	DB (incidents + performance_reviews)
AI Copilot's product dependencies? owners' cities?	Graph multi-hop + DB
Engineers with 2025 H2 perf > 4.0 — what in_progress projects do they own?	DB JOINs
Tell me everything about Frank	KB + DB + Graph
Code review best practices?	Skills
Add "Wednesday deep-work block" to the KB	Ingest (KB write)
Remember I'm on PTO next Monday for a checkup	Memory

8/8 correct on both backends in a verified run:

Backend	Total tool calls	Total time
native (default)	34	194s
sdk + CCR	38	283s

7. Ingest demo — talk to write

In enterprise_demo, the agent has 4 extra tools that write to KB / DB / Graph mid-conversation:

python -m datamind chat
# or open http://127.0.0.1:8000 in a browser

Try:

"add /Users/foo/policy.md to the knowledge base"
  → agent calls kb_add_file, immediately retrievable via kb_search

"import /Users/foo/customers.csv as table customers"
  → agent calls db_import_csv, immediately queryable via db_query_sql

"陈诚 was promoted to Tech Lead, reports to Ann, owns Project Kepler"
  → agent calls graph_add_triples_from_text, LLM extracts triples, graph upserts

Six sample files live in demo-uploads/ — drag any into the browser dropzone to see the full pipeline. See Install §7b and Ingest module.

Full agent in an interactive REPL

python -m datamind chat

╭──── Chat ─────╮
│ DataMind ready · profile=default · model=claude-sonnet-4-6
│ tools=23 · kb_chunks=0 · graph_triples=0 · skills=2
╰───────────────╯
you › How should I run a code review?
ai  › [tool skill_search] {"query":"code review process"}
     [result ok] {"count":1, "results":[{"name":"code-review", …}]}
     [tool skill_get] {"name":"code-review"}
     …

Commands: /new resets history, /exit or Ctrl-D leaves.

HTTP server

python -m uvicorn datamind.server:app --port 8000

Streaming example:

curl -N -X POST localhost:8000/api/chat \
  -H 'Content-Type: application/json' \
  -d '{"message":"Tell me the Status meeting time."}'

data: {"type": "text", "delta": "根据"}
data: {"type": "text", "delta": "公司"}
data: {"type": "tool_use", "name": "kb_search", "input": {"query": "Status meeting time"}, "id": "toolu_…"}
data: {"type": "tool_result", "name": "kb_search", "is_error": false, "preview": "…"}
data: {"type": "text", "delta": "周一"}
…
data: {"type": "done", "iterations": 3, "stop_reason": "end_turn"}

Verify the whole test suite

pytest datamind/tests/
# 95 passed in ~0.6s

No network is required for unit tests — they use in-memory fakes and temp SQLite. Live tests are reserved for the hello_*.py smokes.