RAG Engineer

Overview

This public intake copy packages

plugins/antigravity-awesome-skills-claude/skills/rag-engineer

https://github.com/sickn33/antigravity-awesome-skills

Use it when the operator needs the upstream workflow, support files, and repository context to stay intact while the public validator and private enhancer continue their normal downstream flow.

This intake keeps the copied upstream files intact and uses

metadata.json

ORIGIN.md

RAG Engineer Expert in building Retrieval-Augmented Generation systems. Masters embedding models, vector databases, chunking strategies, and retrieval optimization for LLM applications. Role: RAG Systems Architect I bridge the gap between raw documents and LLM understanding. I know that retrieval quality determines generation quality - garbage in, garbage out. I obsess over chunking boundaries, embedding dimensions, and similarity metrics because they make the difference between helpful and hallucinating. ### Expertise - Embedding model selection and fine-tuning - Vector database architecture and scaling - Chunking strategies for different content types - Retrieval quality optimization - Hybrid search implementation - Re-ranking and filtering strategies - Context window management - Evaluation metrics for retrieval ### Principles - Retrieval quality > Generation quality - fix retrieval first - Chunk size depends on content type and query patterns - Embeddings are not magic - they have blind spots - Always evaluate retrieval separately from generation - Hybrid search beats pure semantic in most cases

Imported source sections that did not map cleanly to the public headings are still preserved below or in the support files. Notable imported sections: Capabilities, Prerequisites, Patterns, Sharp Edges, Limitations.

When to Use This Skill

Use this section as the trigger filter. It should make the activation boundary explicit before the operator loads files, runs commands, or opens a pull request.

• User mentions or implies: building RAG
• User mentions or implies: vector search
• User mentions or implies: embeddings
• User mentions or implies: semantic search
• User mentions or implies: document retrieval
• User mentions or implies: context retrieval

Operating Table

metadata.json

ORIGIN.md

SKILL.md

SKILL.md

## Related Skills

Workflow

This workflow is intentionally editorial and operational at the same time. It keeps the imported source useful to the operator while still satisfying the public intake standards that feed the downstream enhancer flow.

1. Confirm the user goal, the scope of the imported workflow, and whether this skill is still the right router for the task.
2. Read the overview and provenance files before loading any copied upstream support files.
3. Load only the references, examples, prompts, or scripts that materially change the outcome for the current request.
4. Execute the upstream workflow while keeping provenance and source boundaries explicit in the working notes.
5. Validate the result against the upstream expectations and the evidence you can point to in the copied files.
6. Escalate or hand off to a related skill when the work moves out of this imported workflow's center of gravity.
7. Before merge or closure, record what was used, what changed, and what the reviewer still needs to verify.

Imported Workflow Notes

Imported: Capabilities

• Vector embeddings and similarity search
• Document chunking and preprocessing
• Retrieval pipeline design
• Semantic search implementation
• Context window optimization
• Hybrid search (keyword + semantic)

Examples

Example 1: Ask for the upstream workflow directly

Use @rag-engineer to handle <task>. Start from the copied upstream workflow, load only the files that change the outcome, and keep provenance visible in the answer.

Explanation: This is the safest starting point when the operator needs the imported workflow, but not the entire repository.

Example 2: Ask for a provenance-grounded review

Review @rag-engineer against metadata.json and ORIGIN.md, then explain which copied upstream files you would load first and why.

Explanation: Use this before review or troubleshooting when you need a precise, auditable explanation of origin and file selection.

Example 3: Narrow the copied support files before execution

Use @rag-engineer for <task>. Load only the copied references, examples, or scripts that change the outcome, and name the files explicitly before proceeding.

Explanation: This keeps the skill aligned with progressive disclosure instead of loading the whole copied package by default.

Example 4: Build a reviewer packet

Review @rag-engineer using the copied upstream files plus provenance, then summarize any gaps before merge.

Explanation: This is useful when the PR is waiting for human review and you want a repeatable audit packet.

Best Practices

Treat the generated public skill as a reviewable packaging layer around the upstream repository. The goal is to keep provenance explicit and load only the copied source material that materially improves execution.

• Keep the imported skill grounded in the upstream repository; do not invent steps that the source material cannot support.
• Prefer the smallest useful set of support files so the workflow stays auditable and fast to review.
• Keep provenance, source commit, and imported file paths visible in notes and PR descriptions.
• Point directly at the copied upstream files that justify the workflow instead of relying on generic review boilerplate.
• Treat generated examples as scaffolding; adapt them to the concrete task before execution.
• Route to a stronger native skill when architecture, debugging, design, or security concerns become dominant.

Troubleshooting

Problem: The operator skipped the imported context and answered too generically

Symptoms: The result ignores the upstream workflow in

plugins/antigravity-awesome-skills-claude/skills/rag-engineer

metadata.json

ORIGIN.md

Problem: The imported workflow feels incomplete during review

Symptoms: Reviewers can see the generated

SKILL.md

Problem: The task drifted into a different specialization

Symptoms: The imported skill starts in the right place, but the work turns into debugging, architecture, design, security, or release orchestration that a native skill handles better. Solution: Use the related skills section to hand off deliberately. Keep the imported provenance visible so the next skill inherits the right context instead of starting blind.

Related Skills

• @prompt-engineer

  - Use when the work is better handled by that native specialization after this imported skill establishes context.

• @prompt-engineering

  - Use when the work is better handled by that native specialization after this imported skill establishes context.

• @prompt-engineering-patterns

  - Use when the work is better handled by that native specialization after this imported skill establishes context.

• @prompt-library

  - Use when the work is better handled by that native specialization after this imported skill establishes context.

Additional Resources

Use this support matrix and the linked files below as the operator packet for this imported skill. They should reflect real copied source material, not generic scaffolding.

references

references/n/a

examples

examples/n/a

scripts

scripts/n/a

agents

agents/n/a

assets

assets/n/a

Imported Reference Notes

Imported: Prerequisites

• Required skills: LLM fundamentals, Understanding of embeddings, Basic NLP concepts

Imported: Patterns

Semantic Chunking

Chunk by meaning, not arbitrary token counts

When to use: Processing documents with natural sections

• Use sentence boundaries, not token limits
• Detect topic shifts with embedding similarity
• Preserve document structure (headers, paragraphs)
• Include overlap for context continuity
• Add metadata for filtering

Hierarchical Retrieval

Multi-level retrieval for better precision

When to use: Large document collections with varied granularity

• Index at multiple chunk sizes (paragraph, section, document)
• First pass: coarse retrieval for candidates
• Second pass: fine-grained retrieval for precision
• Use parent-child relationships for context

Hybrid Search

Combine semantic and keyword search

When to use: Queries may be keyword-heavy or semantic

• BM25/TF-IDF for keyword matching
• Vector similarity for semantic matching
• Reciprocal Rank Fusion for combining scores
• Weight tuning based on query type

Query Expansion

Expand queries to improve recall

When to use: User queries are short or ambiguous

• Use LLM to generate query variations
• Add synonyms and related terms
• Hypothetical Document Embedding (HyDE)
• Multi-query retrieval with deduplication

Contextual Compression

Compress retrieved context to fit window

When to use: Retrieved chunks exceed context limits

• Extract relevant sentences only
• Use LLM to summarize chunks
• Remove redundant information
• Prioritize by relevance score

Metadata Filtering

Pre-filter by metadata before semantic search

When to use: Documents have structured metadata

• Filter by date, source, category first
• Reduce search space before vector similarity
• Combine metadata filters with semantic scores
• Index metadata for fast filtering

Imported: Sharp Edges

Fixed-size chunking breaks sentences and context

Severity: HIGH

Situation: Using fixed token/character limits for chunking

Symptoms:

• Retrieved chunks feel incomplete or cut off
• Answer quality varies wildly
• High recall but low precision

Why this breaks: Fixed-size chunks split mid-sentence, mid-paragraph, or mid-idea. The resulting embeddings represent incomplete thoughts, leading to poor retrieval quality. Users search for concepts but get fragments.

Recommended fix:

Use semantic chunking that respects document structure:

• Split on sentence/paragraph boundaries
• Use embedding similarity to detect topic shifts
• Include overlap for context continuity
• Preserve headers and document structure as metadata

Pure semantic search without metadata pre-filtering

Severity: MEDIUM

Situation: Only using vector similarity, ignoring metadata

Symptoms:

• Returns outdated information
• Mixes content from wrong sources
• Users can't scope their searches

Why this breaks: Semantic search finds semantically similar content, but not necessarily relevant content. Without metadata filtering, you return old docs when user wants recent, wrong categories, or inapplicable content.

Recommended fix:

Implement hybrid filtering:

• Pre-filter by metadata (date, source, category) before vector search
• Post-filter results by relevance criteria
• Include metadata in the retrieval API
• Allow users to specify filters

Using same embedding model for different content types

Severity: MEDIUM

Situation: One embedding model for code, docs, and structured data

Symptoms:

• Code search returns irrelevant results
• Domain terms not matched properly
• Similar concepts not clustered

Why this breaks: Embedding models are trained on specific content types. Using a text embedding model for code, or a general model for domain-specific content, produces poor similarity matches.

Recommended fix:

Evaluate embeddings per content type:

• Use code-specific embeddings for code (e.g., CodeBERT)
• Consider domain-specific or fine-tuned embeddings
• Benchmark retrieval quality before choosing
• Separate indices for different content types if needed

Using first-stage retrieval results directly

Severity: MEDIUM

Situation: Taking top-K from vector search without reranking

Symptoms:

• Clearly relevant docs not in top results
• Results order seems arbitrary
• Adding more results helps quality

Why this breaks: First-stage retrieval (vector search) optimizes for recall, not precision. The top results by embedding similarity may not be the most relevant for the specific query. Cross-encoder reranking dramatically improves precision for the final results.

Recommended fix:

Add reranking step:

• Retrieve larger candidate set (e.g., top 20-50)
• Rerank with cross-encoder (query-document pairs)
• Return reranked top-K (e.g., top 5)
• Cache reranker for performance

Cramming maximum context into LLM prompt

Severity: MEDIUM

Situation: Using all retrieved context regardless of relevance

Symptoms:

• Answers drift with more context
• LLM ignores key information
• High token costs

Why this breaks: More context isn't always better. Irrelevant context confuses the LLM, increases latency and cost, and can cause the model to ignore the most relevant information. Models have attention limits.

Recommended fix:

Use relevance thresholds:

• Set minimum similarity score cutoff
• Limit context to truly relevant chunks
• Summarize or compress if needed
• Order context by relevance

Not measuring retrieval quality separately from generation

Severity: HIGH

Situation: Only evaluating end-to-end RAG quality

Symptoms:

• Can't diagnose poor RAG performance
• Prompt changes don't help
• Random quality variations

Why this breaks: If answers are wrong, you can't tell if retrieval failed or generation failed. This makes debugging impossible and leads to wrong fixes (tuning prompts when retrieval is the problem).

Recommended fix:

Separate retrieval evaluation:

• Create retrieval test set with relevant docs labeled
• Measure MRR, NDCG, Recall@K for retrieval
• Evaluate generation only on correct retrievals
• Track metrics over time

Not updating embeddings when source documents change

Severity: MEDIUM

Situation: Embeddings generated once, never refreshed

Symptoms:

• Returns outdated information
• References deleted content
• Inconsistent with source

Why this breaks: Documents change but embeddings don't. Users retrieve outdated content or, worse, content that no longer exists. This erodes trust in the system.

Recommended fix:

Implement embedding refresh:

• Track document versions/hashes
• Re-embed on document change
• Handle deleted documents
• Consider TTL for embeddings

Same retrieval strategy for all query types

Severity: MEDIUM

Situation: Using pure semantic search for keyword-heavy queries

Symptoms:

• Exact term searches miss results
• Concept searches too literal
• Users frustrated with both

Why this breaks: Some queries are keyword-oriented (looking for specific terms) while others are semantic (looking for concepts). Pure semantic search fails on exact matches; pure keyword search fails on paraphrases.

Recommended fix:

Implement hybrid search:

• BM25/TF-IDF for keyword matching
• Vector similarity for semantic matching
• Reciprocal Rank Fusion to combine
• Tune weights based on query patterns

Imported: Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.