Chat Configuration

Complete reference for chat system configuration.

Configuration File

Chat is configured via chat.config.yml in the project root:

models:
  plannerModel: gpt-5-mini-2025-08-07
  answerModel: gpt-5-mini-2025-08-07
  embeddingModel: text-embedding-3-small
  reasoning:
    planner: low
    answer: low

tokens:
  planner: 15000
  answer: 15000

retrieval:
  defaultTopK: 8
  maxTopK: 10
  minRelevanceScore: 0.3
  weights:
    textWeight: 0.3
    semanticWeight: 0.5
    recencyLambda: 0.05

moderation:
  input:
    enabled: false
  output:
    enabled: false

cost:
  budgetUsd: 10

Model Configuration

Available Models

| Setting | Description | Default | |---------|-------------|---------| | plannerModel | Model for query planning | gpt-5-mini-2025-08-07 | | answerModel | Model for response generation | gpt-5-mini-2025-08-07 | | embeddingModel | Model for embeddings | text-embedding-3-small |

Reasoning Levels

Control reasoning depth:

| Level | Description | |-------|-------------| | minimal | Fastest, least reasoning | | low | Basic reasoning | | medium | Moderate reasoning | | high | Deep reasoning |

reasoning:
  planner: low    # Query planning reasoning
  answer: low     # Response generation reasoning

Token Limits

Maximum tokens per stage:

tokens:
  planner: 15000  # Query planning budget
  answer: 15000   # Response generation budget

These limits prevent runaway API costs.

Retrieval Configuration

Document Retrieval

retrieval:
  defaultTopK: 8      # Default documents to retrieve
  maxTopK: 10         # Maximum allowed
  minRelevanceScore: 0.3  # Minimum relevance threshold

Scoring Weights

Control how documents are ranked:

weights:
  textWeight: 0.3       # Lexical match importance
  semanticWeight: 0.5   # Semantic similarity importance
  recencyLambda: 0.05   # Time decay factor

textWeight: Higher values favor exact keyword matches.

semanticWeight: Higher values favor semantic similarity.

recencyLambda: Higher values favor newer content. Set to 0 to disable time decay.

Relevance Scoring

Documents with scores below minRelevanceScore (as a fraction of the top result) are excluded:

• 0.3 = Keep docs scoring at least 30% of top match
• 0.5 = Keep docs scoring at least 50% of top match
• 0 = Keep all retrieved documents

Moderation

Input Moderation

Check user messages before processing:

moderation:
  input:
    enabled: true
    model: omni-moderation-latest

Output Moderation

Check generated responses before sending:

moderation:
  output:
    enabled: true
    model: omni-moderation-latest
    refusalMessage: "I can only answer questions about my portfolio."
    refusalBanner: "That request was blocked by my safety filters."

Cost Management

Monthly Budget

cost:
  budgetUsd: 10  # Maximum monthly spend

When budget is reached:

• New requests return 503 error
• CloudWatch alarm triggers
• Email notification sent (if configured)

Disabling Budget

Omit the cost section to disable budget checks:

# cost:
#   budgetUsd: 10

Environment Variables

Required

| Variable | Description | |----------|-------------| | OPENAI_API_KEY | OpenAI API key |

Optional

| Variable | Description | Default | |----------|-------------|---------| | CHAT_DEBUG_LEVEL | Debug verbosity (0-3) | 0 | | OPENAI_COST_METRICS_ENABLED | Enable CloudWatch metrics | false | | OPENAI_COST_METRIC_NAMESPACE | CloudWatch namespace | PortfolioChat/OpenAI | | OPENAI_COST_METRIC_NAME | Metric name | EstimatedCost | | OPENAI_COST_ALERT_EMAIL | Alert email address | - |

Rate Limiting

| Variable | Description | |----------|-------------| | UPSTASH_REDIS_REST_URL | Redis URL for rate limiting | | UPSTASH_REDIS_REST_TOKEN | Redis auth token | | ENABLE_DEV_RATE_LIMIT | Enable rate limiting in dev |

Debug Configuration

Enable verbose logging:

CHAT_DEBUG_LEVEL=3 pnpm dev

| Level | Output | |-------|--------| | 0 | Disabled | | 1 | Basic request/response | | 2 | Include retrieval info | | 3 | Full pipeline debug |

Preprocessing Configuration

Control preprocessing tasks:

CHAT_PREPROCESS_TASKS='profile,persona' pnpm chat:preprocess

Available tasks:

• profile - Portfolio profile
• persona - AI persona
• projects - Project embeddings
• resume - Resume embeddings

Example Configurations

Development

models:
  plannerModel: gpt-4o-mini
  answerModel: gpt-4o-mini
  embeddingModel: text-embedding-3-small

retrieval:
  defaultTopK: 5
  minRelevanceScore: 0.2

moderation:
  input:
    enabled: false
  output:
    enabled: false

cost:
  budgetUsd: 5

Production

models:
  plannerModel: gpt-5-mini-2025-08-07
  answerModel: gpt-5-mini-2025-08-07
  embeddingModel: text-embedding-3-small
  reasoning:
    planner: low
    answer: low

retrieval:
  defaultTopK: 8
  maxTopK: 10
  minRelevanceScore: 0.3
  weights:
    textWeight: 0.3
    semanticWeight: 0.5
    recencyLambda: 0.05

moderation:
  input:
    enabled: true
  output:
    enabled: true
    refusalMessage: "I can only discuss portfolio-related topics."

cost:
  budgetUsd: 10