Ragie RAG

Ragie.ai RAG Skill (OpenClaw Optimized)

This skill enables grounded question answering using Ragie.ai as a RAG backend.

Ragie handles:

• Document chunking
• Embedding
• Vector indexing
• Retrieval
• Optional reranking

The agent handles:

• Deciding when to ingest
• Triggering retrieval
• Constructing grounded prompts
• Producing final answers

Core Principles

1. Never answer without retrieval.
2. Never hallucinate information not present in retrieved chunks.
3. Always cite the document_name when referencing specific facts.
4. If retrieval returns zero relevant chunks, explicitly say:

   "I don't have that information in the current knowledge base."

5. Do not expose API keys or raw API payloads in final answers.

Deterministic Workflow

Case A — User Provides a File or URL

IF the user provides:

• A file
• A document path
• A PDF/URL to ingest

THEN:

1. Execute ingestion:

   python `skills/scripts/ingest.py` --file <path> --name "<document_name>"
   

   OR

   python `skills/scripts/ingest.py` --url "<url>" --name "<document_name>"
   

2. Capture returned document_id.

3. Poll document status:

   python `skills/scripts/manage.py` status --id <document_id>
   

   Repeat until status == ready.

4. Proceed to Retrieval (Case C).

Case B — User Requests Document Management

List documents

python `skills/scripts/manage.py` list

Check document status

python `skills/scripts/manage.py` status --id <document_id>

Delete a document

python `skills/scripts/manage.py` delete --id <document_id>

Return structured results to the user.

Case C — Retrieval (Grounded Question Answering)

Execute:

python `skills/scripts/retrieve.py` \
  --query "<user_question>" \
  --top-k 6 \
  --rerank

Optional flags:

• --partition <name>
• --filter '{"key":"value"}'

Retrieval Output Format

Expected output:

[
  {
    "text": "...",
    "score": 0.87,
    "document_name": "Policy Handbook",
    "document_id": "doc_abc123"
  }
]

Grounded Prompt Construction

After retrieval:

1. Extract all chunk text.
2. Concatenate with separators.
3. Construct this prompt:

SYSTEM:
You are a helpful assistant.
Answer using ONLY the context provided below.
If the context does not contain the answer, say:
"I don't have that information in the current knowledge base."

CONTEXT:
[chunk 1 text]
---
[chunk 2 text]
---
...

USER QUESTION:
{original user question}

4. Generate final answer.
5. Cite document_name when referencing information.

Output Contract

The final response MUST:

• Be grounded only in retrieved chunks
• Cite document_name for factual claims
• Avoid hallucinations
• Avoid mentioning internal execution steps
• Avoid exposing API keys or raw responses
• Clearly state when information is missing

If no chunks are returned:

I don't have that information in the current knowledge base.

API Reference

Base URL:

https://api.ragie.ai

Error Handling

If ingestion fails:

• Report failure clearly.
• Do not proceed to retrieval.

If retrieval fails:

• Retry once.
• If still failing, inform user.

Decision Rules Summary

1. If user uploads content → ingest → wait until ready → retrieve.
2. If user asks question → retrieve immediately.
3. If zero chunks → state knowledge gap.
4. Always use reranking unless explicitly disabled.
5. Never answer without retrieval.

Advanced Usage

• Use metadata filter to narrow retrieval scope.
• Use partitions to separate tenant data.
• Use recency_bias only when time relevance matters.
• Adjust top_k depending on query complexity.

Security

• API keys must be loaded from environment variables.
• .env must not be committed.
• Do not log sensitive headers.

Summary

This skill provides:

• Deterministic ingestion
• Deterministic retrieval
• Strict grounded answering
• Complete Ragie lifecycle management
• Safe and hallucination-resistant RAG execution

End of Skill.