Jeeves

My agentic AI assistant for Open-WebUI. Handles semantic memory, intent classification, workspace ops, and surgical file editing.

What It Does

Jeeves sits between you and your LLM as an intelligent filter. It:

Classifies Intent — Figures out if you're chatting casually, saving info, recalling memories, or asking for a task
Manages Memory — Stores facts, docs, and image descriptions; pulls relevant context automatically
Runs Workspace Ops — Can read, list, and edit files in your workspace
Surgical Edits — Supports append, replace, and insert operations on files

What it can do:

read — View file contents
list — Browse workspace structure
append — Add content to end of files
replace — Find and replace text in files
insert — Insert text at specific positions

Example prompts:

"Show me the readme"
"Insert a credit to me in the readme file"
"Replace 'old text' with 'new text' in config.yaml"
"Add a contributors section to README.md"

Architecture

User message → Jeeves Filter → Intent Classification (Pragmatics)
                    ↓
              ┌─────┴─────┐
              │           │
           task?       recall/save/casual?
              │           │
              ↓           ↓
         Orchestrator   Memory Search
         (reasoning)       │
              │           ↓
              ↓      Context Injected
         Executor API       │
         (file ops)         │
              │             │
              └─────┬───────┘
                    ↓
              LLM Response

Key principle: All task intents go to the Orchestrator for reasoning—no shortcut patterns in the filter.

Services

Service	Port	What it does
`jeeves`	8000	Agent core + semantic memory
`pragmatics_api`	8001	4-class intent classifier (DistilBERT)
`extractor_api`	8002	Image/audio/PDF extraction (GPU)
`orchestrator_api`	8004	Multi-step reasoning engine
`executor_api`	8005	File ops + code execution
`qdrant`	6333	Vector database
`ollama`	11434	Local LLM inference
`open-webui`	8180	Chat UI (filter runs here)

Intent Classification

Pragmatics uses a fine-tuned DistilBERT (98% accuracy) to classify what you want:

Intent	What it means	Example
`casual`	Just chatting, no action	"How are you?"
`save`	Sharing info to remember	"My name is Ian"
`recall`	Asking about past info	"What's my email?"
`task`	Requesting an action	"Add credits to the readme"

File Operations

All task requests go to the Orchestrator, which reasons about what you want and picks the right tool. The Executor then does the actual work.

Executor Tools

Tool	What it does
`read_file`	Read file contents
`write_file`	Overwrite entire file
`replace_in_file`	Find and replace text (surgical)
`insert_in_file`	Insert at start/end/before/after anchor
`append_to_file`	Append to end of file
`list_files`	List directory contents
`scan_workspace`	Recursive search with gitignore, pretty table output
`none`	Skip (change already present)

scan_workspace Output

PATH: /workspace/jeeves
TOTAL: 105 items (27 dirs, 78 files)

NAME                                      TYPE    SIZE      MODIFIED
----------------------------------------------------------------------
filters                                   dir               2025-12-23 03:10:27
layers                                    dir               2025-12-22 23:36:53
README.md                                 file    8.32 KiB  2025-12-23 03:17:10
docker-compose.yaml                       file    6.36 KiB  2025-12-23 03:55:25
...

Running It

Quick Start

# Spin everything up
docker compose up -d --build

# Check status
docker ps

# Tail logs
docker logs jeeves -f
docker logs executor_api -f

Environment Variables

Variable	Default	What it's for
`HOST_WORKSPACE_PATH`	`C:/Code`	Host directory mounted to /workspace
`QDRANT_HOST`	`qdrant`	Vector database host
`OLLAMA_MODEL`	`qwen2.5:14b`	LLM for orchestration
`CLASSIFIER_MODEL`	`distilbert_intent`	Intent classifier model

Filter Sync

The Jeeves filter runs inside Open-WebUI (stored in DB, not mounted). To sync changes:

# Push filter to Open-WebUI (utf-8-sig strips BOM)
$apiKey = (Get-Content "secrets/webui_admin_api_key.txt" -Raw).Trim()
python -c "import requests; f=open('filters/jeeves.filter.py',encoding='utf-8-sig').read(); r=requests.post('http://localhost:8180/api/v1/functions/id/api/update', headers={'Authorization':'Bearer $apiKey'}, json={'id':'api','name':'Jeeves','content':f,'meta':{'toggle':True}}, timeout=10); print(r.status_code)"

Project Structure

jeeves/
├── docker-compose.yaml       # Full stack
├── filters/
│   └── jeeves.filter.py      # Open-WebUI filter (intent routing, workspace ops)
├── layers/
│   ├── memory/               # Semantic memory (port 8000)
│   │   ├── api/memory.py     # /save, /search endpoints
│   │   └── services/         # Embedder, Qdrant, Summarizer
│   ├── pragmatics/           # Intent classifier (port 8001)
│   │   └── services/
│   │       ├── classifier.py       # 4-class DistilBERT
│   │       └── entity_extractor.py # spaCy NER
│   ├── extractor/            # Media extraction (port 8002)
│   │   └── services/         # LLaVA, Whisper, PDF
│   ├── orchestrator/         # Reasoning engine (port 8004)
│   │   └── services/
│   │       ├── reasoning_engine.py # LLM step generation
│   │       ├── code_planner.py     # Specialist for code edits
│   │       └── workspace_state.py  # External state tracking
│   └── executor/             # Code/file execution (port 8005)
│       └── services/
│           ├── file_handler.py      # read, write, replace, insert, append
│           ├── polyglot_handler.py  # Python, Node, PowerShell
│           └── shell_handler.py     # Shell commands
└── .github/
    └── copilot-instructions.md  # AI coding guidelines

How It Works

1. Filter Inlet (jeeves.filter.py)

# You say: "list the files in this workspace"
#
# 1. Intent classified as "task" (99% confidence)
# 2. Task goes to Orchestrator for reasoning
# 3. Orchestrator picks: scan_workspace tool
# 4. Executor runs scan with gitignore support
# 5. Pretty-formatted table injected into context
# 6. LLM presents the results

2. Always Delegate (No Shortcuts)

All task intents go through the Orchestrator:

# Flow in _orchestrate_task():
1. Set workspace context on Orchestrator
2. Get next step (tool + params) from Orchestrator
3. Execute tool via Executor API
4. Return formatted results for context injection

# No hardcoded patterns—Orchestrator decides everything

3. Status Messages

Live feedback while processing:

Icon	What's happening
✨	Thinking
🔍	Scanning workspace
📖	Reading files
✏️	Editing files
⚙️	Running code
💾	Saving to memory
📚	Found memories
✅	Done
❌	Something broke

4. Memory Integration

Save: Extracts facts from the conversation, embeds them, stores in Qdrant
Search: Embeds your query, finds similar memories (cosine > 0.35)
Inject: Prepends relevant memories to LLM context

Quick Test

# Test file append via Executor
Invoke-RestMethod -Uri 'http://localhost:8005/api/execute/tool' -Method Post `
  -ContentType 'application/json' `
  -Body (@{
    tool='append_to_file'
    params=@{path='/workspace/test.txt'; content='Hello from Jeeves'}
    workspace_context=@{workspace_root='/workspace'; cwd='/workspace'; allow_file_write=$true}
  } | ConvertTo-Json -Depth 5)

# Test memory search
Invoke-RestMethod -Uri 'http://localhost:8000/api/memory/search' -Method Post `
  -ContentType 'application/json' `
  -Body (@{user_id='test'; query_text='my name'; top_k=5} | ConvertTo-Json)

# Test intent classification
Invoke-RestMethod -Uri 'http://localhost:8001/api/pragmatics/classify' -Method Post `
  -ContentType 'application/json' `
  -Body (@{text='Add a credit to the readme'} | ConvertTo-Json)


## License

MIT

## Credits

- **Ian Westerfield** - Creator and maintainer

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Repository files navigation

Jeeves

What It Does

Architecture

Services

Intent Classification

File Operations

Executor Tools

scan_workspace Output

Running It

Quick Start

Environment Variables

Filter Sync

Project Structure

How It Works

1. Filter Inlet (jeeves.filter.py)

2. Always Delegate (No Shortcuts)

3. Status Messages

4. Memory Integration

Quick Test

About

Uh oh!

Uh oh!

Contributors 2

Uh oh!

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 29 Commits
.github		.github
filters		filters
layers		layers
ollama		ollama
.gitignore		.gitignore
ARCHITECTURE.md		ARCHITECTURE.md
README.md		README.md
build-base-images.ps1		build-base-images.ps1
docker-compose.yaml		docker-compose.yaml

ianwesterfield/jeeves

Folders and files

Latest commit

History

Repository files navigation

Jeeves

What It Does

Architecture

Services

Intent Classification

File Operations

Executor Tools

scan_workspace Output

Running It

Quick Start

Environment Variables

Filter Sync

Project Structure

How It Works

1. Filter Inlet (jeeves.filter.py)

2. Always Delegate (No Shortcuts)

3. Status Messages

4. Memory Integration

Quick Test

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Uh oh!

Contributors 2

Uh oh!

Languages