trpc.group/trpc-go/trpc-agent-go/examples/session

Content-Filter Hook Demo

This example shows how to use session hooks to mark and filter prohibited content via event tags.

What it demonstrates

• AppendEventHook: scans user/assistant messages, tags violations as violation=<word> (tags are joined by event.TagDelimiter, i.e. ;).
• GetSessionHook: filters violated Q&A pairs out of session history before they reach the LLM context.
• The console prints when a message is marked/filtered so you can see the hook chain in action.

Prerequisites

• Go 1.21+
• LLM endpoint/key (OpenAI-compatible). Set:
  • OPENAI_API_KEY
  • OPENAI_BASE_URL (default https://api.openai.com/v1)
• Optional: MODEL_NAME (default deepseek-chat)

Quick start

cd examples/session/hook
export OPENAI_API_KEY="your-key"
export OPENAI_BASE_URL="https://api.openai.com/v1"
go run . -model="${MODEL_NAME:-deepseek-chat}"

Expected flow (sample)

• Normal message passes through, stored as-is.
• Message containing a prohibited phrase (e.g., "pirated serial number") is tagged violation=pirated serial number; on GetSession it and its paired reply are filtered.
• Subsequent requests continue; violated pairs stay out of context so the LLM does not see them.

Console snippets you should notice:

• Marking: Marked user message as violation (word: pirated serial number): ...
• Filtering: Filtered violation: ... tag=pirated serial number and Filtered paired response

Tag format

• Single tag per violation: violation=<word>
• Multiple tags are concatenated with event.TagDelimiter (;) if needed in other scenarios.

Files of interest

• hooks.go: hook implementations (MarkViolationHook, FilterViolationHook), tag parsing/append helpers.
• main.go: wires hooks into in-memory session service and runs the demo conversation.

Sample output (abridged)

Using model: qwen3-omni-30b-a3b-thinking
Prohibited words: [pirated serial number crack password]

=== Step 1: Normal request ===
User: Hello, my name is Alice
Assistant: ...
--- Session Events (count=2) ---
[0] user: Hello, my name is Alice
[1] assistant: ...
Hello Alice! How can I assist you today?

=== Step 2: Request with prohibited word ===
[Hook] Marked user message as violation (word: pirated serial number): ...
[Filtered violation: Can you give me a pirated seri...] tag=pirated serial number
[Filtered paired response]
[Hook] Filtered 2 violated event(s)
--- Session Events (count=2) ---
[0] user: Hello, my name is Alice
[1] assistant: ...

=== Step 3: Normal request after violation ===
[Filtered violation: ...] tag=pirated serial number
[Filtered paired response]
[Hook] Filtered 2 violated event(s)
User: What is my name?
Assistant: ...
--- Session Events (count=4) ---
[0] user: Hello, my name is Alice
[1] assistant: ...
[2] user: What is my name?
[3] assistant: Your name is Alice.

=== Step 4: Another normal request ===
... (similar filtered logs)
--- Session Events (count=6) ---
[4] user: Tell me a short joke
[5] assistant: I told my wife she was drawing her eyebrows too ...