Description

@people1/n8n-nodes-bedrock-advanced

People1 fork of n8n-nodes-bedrock-advanced@0.5.2 by Amir Souchami.

Why this fork?

The original node extracts Bedrock prompt cache metrics (cacheReadInputTokens, cacheWriteInputTokens) from the API response but does not propagate them to the N8N node output. They are only written to container debug logs, which requires admin SSH access to read.

This fork fixes that: cache metrics now appear in the standard N8N execution output and can be queried via the REST API (GET /api/v1/executions/{id}) by any developer.

Coexistence with the original

This fork uses different internal node type names so it can be installed alongside the original:

| | Original | Fork (P1) |
|—|—|—|
| Converse API node | AWS Bedrock Chat Model (Advanced) | AWS Bedrock Chat Model (Advanced P1) |
| InvokeModel node | Bedrock Claude | Bedrock Claude (P1) |
| Internal type | lmChatAwsBedrockAdvanced | lmChatAwsBedrockAdvancedP1 |
| Internal type | lmChatBedrockClaude | lmChatBedrockClaudeP1 |

Both appear in the N8N node picker. Developers choose which to use per workflow.

What changed

Patches applied to both nodes:

1. Added cache fields to llmOutput.tokenUsage — so cache metrics travel through LangChain’s callback chain
2. Custom tokensUsageParser for N8nLlmTracing — so N8N’s tracing callback preserves cache fields instead of stripping them
3. Streaming path fix (BedrockClaude only) — reads cache metrics from response_metadata.usage for the streaming aggregation path
4. Renamed node types — P1 suffix to coexist with the original package

Output format (after fix)

{
  "tokenUsage": {
    "completionTokens": 500,
    "promptTokens": 150,
    "totalTokens": 650,
    "cacheReadInputTokens": 120,
    "cacheWriteInputTokens": 0
  }
}

When cache is disabled or not applicable, cacheReadInputTokens and cacheWriteInputTokens are 0.

Multi-cachepoint system prompt (v0.6.0+)

The AWS Bedrock Chat Model (Advanced P1) node exposes an option
systemPromptBlocks that lets the system prompt be split into multiple
blocks with a cachePoint marker between each pair.

The Bedrock Converse API supports up to 4 cachepoints per request. With
this option you can keep the fixed part of the prompt (rules, identity,
protocols) cached across requests while only the variable part (user
context, role data, session state) pays the cache-write cost when it
changes.

Usage

1. Produce the blocks upstream in a Code node (sibling of the AI Agent):

   return [{
     json: {
       systemBlocks: [
         "REGLA #1...nIdentidad...nProtocolos...",   // fixed block
         "Rol del usuario X...nTools disponibles...", // variable block
       ],
     },
   }];
   

2. Wire the array to the Bedrock node option (expression):

   {{ $('Generate Prompt Blocks').item.json.systemBlocks }}
   

3. Leave the AI Agent systemMessage empty. When systemPromptBlocks
is set, it REPLACES the system message content. The node emits a
warn log if the Agent’s system message has content.

Rules

• Each block should be ≥ ~1024 tokens (~4000 chars) to reach Bedrock’s

minimum cacheable size; shorter blocks produce a warn log but do not
fail (Bedrock just ignores the cachepoint).

• Max 4 cachepoints per request; blocks beyond the 4th are merged into

the 4th with a nn separator + error log.

• Ignored if cacheSystemPrompt: false or enablePromptCaching: false.
• Accepts either an array literal or a JSON string that parses to one.

Full contract: docDevsPeople1/planesClaude/CACHINGREFACTORCONTRACT.md §5.

Installation

On the N8N server (admin only)

1. Find the N8N main container
docker ps --format '{{.Names}}' | grep n8n
2. Install the fork alongside the original (replace CONTAINER)
docker exec CONTAINER sh -c "cd /home/node/.n8n/nodes && npm install github:franlealp1/n8n-nodes-bedrock-advanced-p1"
3. Restart N8N to pick up the new nodes
docker restart CONTAINER

Both the original and P1 nodes will appear in the node picker.

Persisting across Coolify deploys

Add to the N8N service’s Custom Start Command in Coolify:

cd /home/node/.n8n/nodes && npm install github:franlealp1/n8n-nodes-bedrock-advanced-p1 && cd / && n8n start

Or in a post-deploy script if available.

Querying cache metrics (developers)

After installation, any developer with N8N API access can query cache metrics:

Get execution details (replace ID with execution ID)
curl -s "$N8NBASEURL/api/v1/executions/724002" 
  -H "X-N8N-API-KEY: $N8NAPIKEY" | 
  python3 -c "
import json, sys
data = json.load(sys.stdin)
for nodename, nodedata in data.get('data', {}).get('resultData', {}).get('runData', {}).items():
    for run in node_data:
        for output in run.get('data', {}).get('ai_languageModel', []):
            for item in output:
                tu = item.get('json', {}).get('tokenUsage', {})
                if tu.get('cacheReadInputTokens', 0) > 0 or tu.get('cacheWriteInputTokens', 0) > 0:
                    print(f'{node_name}: read={tu["cacheReadInputTokens"]}, write={tu["cacheWriteInputTokens"]}')
"

Original features (unchanged)

All features from the original node are preserved:

• AWS Bedrock Chat Model (Advanced P1) — Converse API, multi-model, prompt caching (system/tools/history), configurable TTL
• Bedrock Claude (P1) — InvokeModel API, Claude-specific features (web search, computer use, bash, text editor, tool search, programmatic tool calling, 1M context, context compaction)

License

MIT (same as original)