from exa_py import Exaexa = Exa(api_key="your-api-key")# Semantic search — describe what you want in plain languageresult = exa.search(    "Why did NVIDIA stock drop despite strong Q4 2026 earnings",    type="neural",          # semantic search, not keyword    num_results=10,    start_published_date="2026-02-25",   # only search for latest information    contents={        "text": {"max_characters": 3000},       # get full article text        "highlights": {"num_sentences": 3},     # key sentences        "summary": {"query": "What caused the stock drop?"}  # AI summary    })for r in result.results:    print(f"[{r.published_date}] {r.title}")    print(f"  Summary: {r.summary}")    print(f"  URL: {r.url}\n")

# Only financial reports from trusted sourcesearnings = exa.search(    "NVIDIA Q4 2026 earnings analysis",    category="financial report",    num_results=5,    include_domains=["reuters.com", "bloomberg.com", "wsj.com"],    contents={"highlights": True})

# "Show me more analysis like this one"similar = exa.find_similar(    url="https://fortune.com/2026/02/25/nvidia-nvda-earnings-q4-results",    num_results=10,    start_published_date="2026-02-20",    contents={"text": {"max_characters": 2000}})

# Complex question — needs multi-source synthesisdeep_result = exa.search(    "How will Middle East tensions affect global tech supply chain and semiconductor stocks",    type="deep",    num_results=8,    contents={        "summary": {            "query": "Extract: 1) supply chain risk 2) stock impact 3) timeline"        }    })

# Breaking news only — today' iss date 2026-03-05breaking = exa.search(    "US stock market breaking news today",    category="news",    num_results=20,    start_published_date="2026-03-05",    contents={"highlights": {"num_sentences": 2}})

from pymilvus import MilvusClient, DataTypefrom openai import OpenAImilvus = MilvusClient("./my_investment_brain.db")llm = OpenAI()def embed(text: str) -> list[float]:    return llm.embeddings.create(        input=text, model="text-embedding-3-small"    ).data[0].embedding# Collection 1: past decisions and lessons# Every trade I make, I write a short review afterwardmilvus.create_collection(    "decisions",    dimension=1536,    auto_id=True)# Collection 2: my preferences and biases# Things like "I tend to hold tech stocks too long"milvus.create_collection(    "preferences",    dimension=1536,    auto_id=True)# Collection 3: market patterns I've observed# "When VIX > 30 and Fed is dovish, buy the dip usually works"milvus.create_collection(    "patterns",    dimension=1536,    auto_id=True)

import jsondef extract_and_store_memories(conversation: list[dict]) -> int:    """    After each chat session, extract personal insights    and store them in Milvus automatically.    """    # Ask LLM to extract structured memories from conversation    extraction_prompt = """    Analyze this conversation and extract any personal investment insights.    Look for:    1. DECISIONS: specific buy/sell actions and reasoning    2. PREFERENCES: risk tolerance, sector biases, holding patterns    3. PATTERNS: market observations, correlations the user noticed    4. LESSONS: things the user learned or mistakes they reflected on    Return a JSON array. Each item has:    - "type": one of "decision", "preference", "pattern", "lesson"    - "content": the insight in 2-3 sentences    - "confidence": how explicitly the user stated this (high/medium/low)    Only extract what the user clearly expressed. Do not infer or guess.    If nothing relevant, return an empty array.    """    response = llm.chat.completions.create(        model="gpt-4o",        messages=[            {"role": "system", "content": extraction_prompt},            *conversation        ],        response_format={"type": "json_object"}    )    memories = json.loads(response.choices[0].message.content)    stored = 0    for mem in memories.get("items", []):        if mem["confidence"] == "low":            continue    # skip uncertain inferences        collection = {            "decision": "decisions",            "lesson": "decisions",            "preference": "preferences",            "pattern": "patterns"        }.get(mem["type"], "decisions")        # Check for duplicates — don't store the same insight twice        existing = milvus.search(            collection,            data=[embed(mem["content"])],            limit=1,            output_fields=["text"]        )        if existing[0] and existing[0][0]["distance"] > 0.92:            continue    # too similar to existing memory, skip        milvus.insert(collection, [{            "vector": embed(mem["content"]),            "text": mem["content"],            "type": mem["type"],            "source": "chat_extraction",            "date": "2026-03-05"        }])        stored += 1    return stored

def recall_my_experience(situation: str) -> dict:    """    Given a current market situation, retrieve my relevant    past experiences, preferences, and observed patterns.    """    query_vec = embed(situation)    # Search all three collections in parallel    past_decisions = milvus.search(        "decisions", data=[query_vec], limit=3,        output_fields=["text", "date", "tag"]    )    my_preferences = milvus.search(        "preferences", data=[query_vec], limit=2,        output_fields=["text", "type"]    )    my_patterns = milvus.search(        "patterns", data=[query_vec], limit=2,        output_fields=["text"]    )    return {        "past_decisions": [h["entity"] for h in past_decisions[0]],        "preferences": [h["entity"] for h in my_preferences[0]],        "patterns": [h["entity"] for h in my_patterns[0]]    }# When Agent analyzes current tech selloff:context = recall_my_experience(    "tech stocks dropping 3-4% due to Middle East tensions, March 2026")# context now contains:# - My 2024-10 lesson about not panic-selling during ME crisis# - My preference: "I tend to overweight geopolitical risk"# - My pattern: "tech selloffs from geopolitics recover in 1-3 weeks"

---name: post-earnings-evaldescription: >  Evaluate whether to buy, hold, or sell after an earnings report.  Trigger when discussing any stock's post-earnings price action,  or when a watchlist stock reports earnings.---## Post-Earnings Evaluation FrameworkWhen analyzing a stock after earnings release:### Step 1: Get the factsUse Exa to search for:- Actual vs expected: revenue, EPS, forward guidance- Analyst reactions from top-tier sources- Options market implied move vs actual move### Step 2: Check my historyUse Milvus recall to find:- Have I traded this stock after earnings before?- What did I get right or wrong last time?- Do I have a known bias about this sector?### Step 3: Apply my rules- If revenue beat > 5% AND guidance raised → lean BUY- If stock drops > 5% on a beat → likely sentiment/macro driven  - Check: is the drop about THIS company or the whole market?  - Check my history: did I overreact to similar drops before?- If P/E > 2x sector average after beat → caution, priced for perfection### Step 4: Output formatSignal: BUY / HOLD / SELL / WAITConfidence: High / Medium / LowReasoning: 3 bullets maxPast mistake reminder: what I got wrong in similar situationsIMPORTANT: Always surface my past mistakes. I have a tendency tolet fear override data. If my Milvus history shows I regrettedselling after a dip, say so explicitly.

# HEARTBEAT.md — runs every 30 minutes automatically## Morning brief (6:30-7:30 AM only)- Use Exa to search overnight US market news, Asian markets, oil prices- Search Milvus for my current positions and relevant past experiences- Generate a personalized morning brief (under 500 words)- Flag anything related to my past mistakes or current holdings- End with 1-3 action items- Send the brief to my phone## Price alerts (during US market hours 9:30 AM - 4:00 PM ET)- Check price changes for: NVDA, TSM, MSFT, AAPL, GOOGL- If any stock moved more than 3% since last check:  - Search Milvus for: why I hold this stock, my exit criteria  - Generate alert with context and recommendation  - Send alert to my phone## End of day summary (after 4:30 PM ET on weekdays)- Summarize today's market action for my watchlist- Compare actual moves with my morning expectations- Note any new patterns worth remembering

作者介绍

张晨

Zilliz Algorithm Engineer