recall

A self-contained Windows CLI that lets you chat with your personal corpus of files using natural language. All data stays on your machine — no cloud calls during operation.

╔══════════════════════════════════════════════╗
║  recall — your personal memory               ║
║  KB: 42 files · 1 204 chunks                 ║
║  Type /help for commands                     ║
╚══════════════════════════════════════════════╝

---

Prerequisites

Requirement	Minimum
Windows	10 x64 or later
Ollama	Latest stable
Everything	1.4+ (optional — improves search speed)

.NET 8 is bundled — a separate runtime install is not required.

---

Supported File Types

recall uses two extraction methods, applied in order. Only plain text is extracted — no images or embedded media are processed.

Tier 1 — Open XML (built-in, always works)

These formats are ZIP-based archives and are extracted using .NET's built-in compression library. No extra software required.

Extension	Format
.docx	Microsoft Word (2007 and later)
.xlsx	Microsoft Excel (2007 and later)
.pptx	Microsoft PowerPoint (2007 and later)
.txt .md .csv .log	Plain text variants

Tier 2 — Windows IFilter (requires registered COM handler)

IFilter is a Windows COM interface that extracts text from arbitrary file types. It works when the corresponding IFilter handler is installed and registered on your machine. The handler is usually installed alongside the application that owns the format.

Extension	Format	Typical handler
.pdf	Adobe PDF	Adobe Acrobat / Windows PDF viewer
.msg	Outlook email	Microsoft Office
.doc .xls .ppt	Legacy Office (pre-2007 binary)	Microsoft Office
.eml	Email (MIME)	Windows Search built-in
.html .htm	HTML	Windows Search built-in
.rtf	Rich Text Format	Windows Search built-in
Any other extension	—	Whatever IFilter is registered for that type

OneDrive users: IFilter opens the file at the OS level, which can trigger a network sync for cloud-only files. recall handles this gracefully — if IFilter doesn't respond within 15 seconds the file is skipped with a log message. Ensure files are locally available for best results.

Click-to-Run Office: The Office IFilter shipped with Click-to-Run installations may not respond to out-of-process COM calls. In that case recall automatically falls back to the Open XML extractor for .docx, .xlsx, and .pptx files.

Not supported

• .doc, .xls, .ppt (legacy binary Office) when no IFilter is installed
• .pdf when no PDF IFilter is installed (e.g. no Acrobat or Windows PDF handler)
• Binary files: images, audio, video, executables, archives
• Password-protected documents

---

Native Libraries

Both DLLs are bundled in the release zip inside the libs\ folder — you do not need to download them separately.

DLL	Purpose
libs\vec0.dll	Vector similarity search — sqlite-vec
libs\Everything64.dll	Fast file discovery — Everything SDK

Everything64.dll requires the Everything application to be installed and running. Without it, recall falls back to Windows Desktop Search (WDS), which is always available on Windows 10/11 but may not index every folder.

---

Supporting Software

Software	URL	Notes
Ollama	https://ollama.com/download/windows	Required. Runs all LLMs locally.
nomic-embed-text	https://ollama.com/library/nomic-embed-text	Required. Embedding model (ollama pull nomic-embed-text).
qwen3:14b	https://ollama.com/library/qwen3	Default chat model (ollama pull qwen3:14b).
Everything	https://www.voidtools.com	Optional. Enables faster file discovery than WDS.
sqlite-vec	https://github.com/asg017/sqlite-vec/releases	Bundled. Source of vec0.dll.
Everything SDK	https://www.voidtools.com/downloads/	Bundled. Source of Everything64.dll.

The installer (install.ps1) handles Ollama and model pulls automatically. The DLLs are pre-bundled — no manual downloads required.

---

Installation

# Run once from the recall directory
.\install.ps1

The installer will:

1. Check for Ollama and download it if missing
2. Pull the embedding model (nomic-embed-text)
3. Pull the default chat model (qwen3:14b — see appsettings.json)
4. Verify the bundled libs\vec0.dll is present
5. Copy files to %APPDATA%\recall\ (including libs\) and add it to your PATH

---

First Run

recall

On first launch recall will:

• Create %APPDATA%\recall\recall.db automatically
• Check that Ollama is running and both models are available
• Show the startup banner with an empty KB (0 files, 0 chunks)
• Prompt you to run /setup if no search paths are configured

Configure which folders to search:

recall> /setup

Then ingest your first files:

recall> /ingest quarterly report

recall will search for matching files, show you the candidates, and ask whether to ingest them.

---

Command Reference

Command	Description
/setup	Configure which folders recall searches
/help	Show command list
/search <query>	Search for files without ingesting
/ingest <query>	Find files and add new/stale ones to the KB
/kb	Show KB stats (files, chunks, last updated)
/clear	Clear the current conversation context
/forget <path>	Remove a file and its chunks from the KB
/truncate	Delete all files and chunks from the KB (prompts for confirmation)
/exit or /quit	Exit recall
anything else	Chat query against your knowledge base

Colour coding in /search and /ingest

Colour	Meaning
Green	File is already in the knowledge base and up to date
Yellow	File is in the KB but has been modified — will be re-ingested
Dim	File has not been ingested yet

---

Privacy

• All embeddings and chat inference run locally via Ollama
• No data is sent to any external service during normal operation
• The only network calls are Ollama model downloads (ollama pull …) during setup
• recall.db lives in %APPDATA%\recall\ and is never transmitted anywhere

---

Build from Source

Requires .NET 8 SDK.

dotnet publish Recall.Cli/Recall.Cli.csproj `
  -c Release `
  -r win-x64 `
  --self-contained true `
  -p:PublishSingleFile=false `
  -o ./publish

The post-publish step copies libs\ and appsettings.json into ./publish automatically.

---

Models

Role	Default	Notes
Embeddings	nomic-embed-text (768-dim)	Required. Do not change — chunk vectors in the DB are tied to this model's dimensions.
Chat	qwen3:14b	Change freely. Any Ollama model works.

To use a different chat model, edit appsettings.json → Ollama.ChatModel and restart.

# Example: switch to a smaller model
ollama pull qwen3:8b
# then set "ChatModel": "qwen3:8b" in appsettings.json

If you change the embedding model, delete recall.db and re-ingest all files — embeddings from different models are not compatible.

---

Configuration (appsettings.json)

{
  "Recall": {
    "DbPath": "%APPDATA%\\recall\\recall.db",
    "Vec0DllPath": "libs\\vec0.dll",
    "EverythingDllPath": "libs\\Everything64.dll",
    "SearchPaths": []
  },
  "Ollama": {
    "BaseUrl": "http://localhost:11434",
    "EmbeddingModel": "nomic-embed-text",
    "ChatModel": "qwen3:14b",
    "ChatContextWindow": 8192
  },
  "Ingestion": {
    "ChunkSize": 512,
    "ChunkOverlap": 100,
    "MaxExtractedCharsPerFile": 500000
  },
  "Retrieval": {
    "TopK": 5,
    "MaxDistance": 1.0,
    "MaxContextChars": 6000
  }
}

SearchPaths is populated by /setup. %APPDATA% and %USERPROFILE% are expanded at runtime everywhere.

---

Author

James S. K. Makumbi