Skip to main content
Mako includes a lightweight retrieval-augmented generation (RAG) system that gives the model curated knowledge about the Base ecosystem — without relying on the model’s training data alone.

​
How it works

  User message
       │
       ▼
  ┌────────────────┐     ┌─────────────────────┐
  │  Keyword scan   │────▶│  Knowledge entries   │
  │  (85+ triggers) │     │  (15 JSON files)     │
  └────────────────┘     └──────────┬────────────┘
                                    │
                              Top 2 matches
                              (~500 tokens)
                                    │
                                    ▼
                         ┌─────────────────────┐
                         │  Injected as system  │
                         │  message before      │
                         │  inference           │
                         └─────────────────────┘
Before every inference call, the gateway scans the user’s message for known project keywords. If a match is found, relevant context is injected as a system message — so the model has accurate, up-to-date information without needing a separate retrieval step.

​
Two retrieval modes

Auto-injection

Keywords in the user’s message trigger context injection automatically, before the model sees the message. Zero latency.

Explicit search

The model can call the knowledge_search tool directly to query the knowledge base on demand.

​
Auto-injection

The getContextForMessage() function runs on every request:
  1. Scans the user’s message against 85+ keyword triggers built from each project’s keywords array and name
  2. Scores matches by keyword hit count
  3. Injects the top 2 project summaries (~500 tokens) as a system message

​
Keyword matching rules

RuleBehavior
Word boundaryKeywords are matched as whole words using regex word boundaries to avoid false positives (e.g., “base” won’t match “database”)
Short keywords (≤3 chars)Require a $ prefix or strict word-boundary matching to prevent noise (except common tickers like eth, bnb, uni)
Token symbols$AERO, $VIRTUAL, etc. are matched with the $ prefix
ScoringEach keyword hit adds to the entry’s score. Top 2 entries by score are injected
TruncationProject details are capped at 600 characters to conserve tokens

​
Injected format

[Knowledge Context]
Aerodrome (DeFi): The central DEX and liquidity hub on Base.
Aerodrome is a ve(3,3) DEX that serves as the primary...
Token: $AERO on base
Website: https://aerodrome.finance
The model can also call knowledge_search directly when it needs structured information. This uses a weighted scoring system:
FieldScore weight
Project name+10
Keywords+5 per match
Category+3
Summary+2
Details+1
Returns up to 5 results, each with name, category, summary, details, token, and website. If called with an empty query, it returns all entries (name, category, and summary only) — useful for browsing.

​
Knowledge entries

Each entry is a JSON file in gateway/src/knowledge/. The gateway loads all entries at startup.

​
Entry schema

{
  "name": "Aerodrome",
  "category": "DeFi",
  "summary": "The central DEX and liquidity hub on Base.",
  "keywords": ["aerodrome", "aero", "dex", "liquidity", "ve33"],
  "token": {
    "symbol": "AERO",
    "chain": "base"
  },
  "links": {
    "website": "https://aerodrome.finance",
    "docs": "https://docs.aerodrome.finance"
  },
  "details": "Aerodrome is a ve(3,3) DEX...",
  "notable_agents": [
    {
      "name": "AeroBot",
      "description": "Automated LP management"
    }
  ]
}

​
Current entries

The knowledge base ships with 15 curated entries:
ProjectCategoryToken
AerodromeDeFi$AERO
MoonwellDeFi$WELL
MorphoDeFi$MORPHO
Aave (Base)DeFi$AAVE
Uniswap (Base)DeFi$UNI
Virtuals ProtocolAI Agents$VIRTUAL
Venice AIAI Agents—
ClankerAI Agents—
AIXBTAI Agents$AIXBT
LunaAI Agents$LUNA
BankrAI Agents$BANKR
FarcasterSocial—
ZoraSocial—
Base ChainInfrastructure—
MakoInfrastructure$MAKO

​
Adding a new entry

Create a JSON file in gateway/src/knowledge/ following the schema above. The gateway loads all .json files from this directory at startup — no code changes required, just restart the server.
Choose keywords carefully. Each keyword becomes a trigger for auto-injection, so avoid overly generic terms that would match irrelevant queries.