synthesis-llm-setup

LLM Setup

Configure Claude Projects, ChatGPT GPTs, Gemini Gems, and other LLM platforms using compiled AI Knowledge content from the ragbot system.

Configuration

These values are user-specific. Update them for your environment.

Setting	Value	Description
ai_knowledge_repo	ai-knowledge-{name}/	Your ai-knowledge repository root
compiled_instructions_path	compiled/{project}/instructions/	Path to LLM-specific compiled instructions
knowledge_file	all-knowledge.md	Concatenated knowledge file (auto-generated by CI/CD)
source_dir	source/	Directory to edit directly (knowledge concatenation is automatic)

Architecture overview

The AI Knowledge system produces platform-specific content for each LLM's project/custom instruction system.

Operation	Where	When
Knowledge concatenation (all-knowledge.md)	CI/CD (GitHub Actions)	Every push to source/
Instruction compilation	Local (ragbot compile)	When instructions change (rare)
RAG indexing	Local (ragbot index)	When content changes and RAG is needed

Output structure

ai-knowledge-{name}/
├── compiled/
│   └── {project}/
│       └── instructions/         # LLM-specific custom instructions
│           ├── claude.md
│           ├── chatgpt.md
│           └── gemini.md
└── all-knowledge.md              # Concatenated knowledge (auto-generated by CI/CD)

Key principle: Edit source/ files directly. Knowledge concatenation is automatic.

Knowledge delivery strategy:

• Claude: GitHub sync the repo (Claude reads all-knowledge.md directly)
• ChatGPT: Upload all-knowledge.md
• Gemini: Upload all-knowledge.md (works within the 10-file Gem limit)

Claude Projects setup

Custom instructions

1. Create a new Claude Project (or open an existing one).
2. Go to Project Knowledge, then Custom Instructions.
3. Copy content from compiled/{project}/instructions/claude.md.
4. Paste into the custom instructions field.

Knowledge files

Option A: GitHub sync (recommended)

1. Connect your GitHub account to Claude.
2. Sync the repository containing your ai-knowledge repo.
3. Claude indexes all-knowledge.md and source files automatically.

Option B: Manual upload

1. Go to Project Knowledge, then Files.
2. Upload all-knowledge.md from the repo root.

ChatGPT GPT setup

Creating a GPT

1. Go to https://chat.openai.com/gpts/editor
2. Click "Create a GPT".
3. Configure:
   • Name: Your project name
   • Description: Brief description
   • Instructions: Copy from compiled/{project}/instructions/chatgpt.md

Knowledge files

1. In the GPT editor, go to the Knowledge section.
2. Upload all-knowledge.md from the repo root.

Gemini Gems setup

Creating a Gem

1. Go to https://gemini.google.com/gems
2. Create a new Gem.
3. Paste instructions from compiled/{project}/instructions/gemini.md.

Knowledge files

1. Upload all-knowledge.md from the repo root.
2. This single file contains all runbooks and datasets merged together.
3. Works well within Gemini's 10-file limit per Gem.

Other LLMs (Grok, etc.)

1. Copy instructions from compiled/{project}/instructions/ (use the closest match).
2. Upload all-knowledge.md from the repo root.

Compilation and inheritance

How it works

Each user compiles projects in their own repo. What content gets included depends on inheritance.

Example: Compiling in ai-knowledge-personal:

compiled/
├── personal/                     # Baseline (ragbot + personal)
├── company/                      # personal + company merged
├── client-a/                     # personal + company + client-a merged
└── client-b/                     # personal + client-b merged

Example: Compiling in ai-knowledge-company (team member without access to personal):

compiled/
├── company/                      # Baseline (ragbot + company, NO personal)
├── client-a/                     # company + client-a (NO personal)
└── client-c/                     # company + client-c

Privacy model

Content is only included if the user has access to the source repo:

• Private content (ai-knowledge-{personal}) only appears in that user's compilations
• Team members get team content but not personal content
• Clients only get client-specific content

Running instruction compilation

# Compile instructions for a project
ragbot compile --project {name}

# Without LLM API calls (just assemble)
ragbot compile --project {name} --no-llm

# Force recompile (ignore cache)
ragbot compile --project {name} --force

# Verbose output
ragbot compile --project {name} --verbose

Knowledge concatenation (all-knowledge.md) is handled automatically by CI/CD -- no manual step needed.

Step-by-step workflow

1. Compile instructions (only if instructions changed):

   ragbot compile --project {name} --no-llm
   

2. For each project (e.g., client-a):

   • Claude: Create project, copy compiled/client-a/instructions/claude.md to custom instructions, GitHub sync the repo
   • ChatGPT: Create GPT, copy compiled/client-a/instructions/chatgpt.md to instructions, upload all-knowledge.md
   • Gemini: Create Gem, copy compiled/client-a/instructions/gemini.md to instructions, upload all-knowledge.md

3. Verify by testing each project with a representative query.

Updating projects

When to update

Knowledge (all-knowledge.md): Updates automatically via CI/CD on every push to source/. No manual step.

Instructions: Recompile only when instructions change (rare):

ragbot compile --project {name} --force

LLM sync:

• Claude: GitHub sync auto-updates
• ChatGPT/Gemini: Re-upload all-knowledge.md after source changes

Troubleshooting

"Instructions too long"

• Move detailed content to knowledge files
• Check manifest.yaml for token counts
• Keep instructions focused on identity and behavior

"Knowledge not being used"

• Verify all-knowledge.md was uploaded correctly
• Check if content is in instructions vs knowledge
• For Claude: ensure GitHub sync is active and pointing to the repo

"Inheritance not working"

• Verify my-projects.yaml exists in the personal repo
• Check inheritance chain in compile-config.yaml
• Run with --verbose to see inheritance resolution

"Content from wrong repo appearing"

• Check which repo you are compiling in
• Verify you have access to expected repos
• Remember: content only comes from repos you can access