An intelligent file indexer with powerful AI search (RAG engine), automatic OCR, and a seamless MCP interface.

• Unlock your documents with semantic AI search & query
• Files are split using semantic chunking with context headers and committed to a local database.
• RAG engine¹ uses reranking and expanding of retrieved chunks

¹ Retrieval Augmented Generation is the method of matching pre-made snippets of information to a query.

• Includes local AI file system indexer
• Natively ingests PDFs, images, Markdown, plaintext, and more…
• Selects and tracks files using patterns like ~/Documents/*.pdf
• Transcribes images using automatic OCR (experimental) and entity extraction
• Changes are automatically synced to a local Qdrant vector database.

• Supports many AI providers and MCP
• OpenAI or compatible API ¹ for best performance
• Ollama and LM Studio for best privacy (local LLM)
• Integrates with your workflow** via a built-in MCP server.

¹ Includes xAI / Grok and Claude OpenAI compatible APIs. Simply adjust the URL settings and overwrite OPENAI_API_KEY.

• Fully resumable parallel processing
• Processes multiple files at once using optimized multi-threading.
• Uses AI cache and generous request retry logic for all network requests.
• Leverages AI structured output with high-quality prompts and schemas.

(If you can't see the diagram below, view it on Mermaid.live)

graph LR

  %% Ingestion Pipeline
  subgraph Ingestion
    A[<b>Track and Commit Files</b><br>Supports Profiles] --> B[<b>Ingest Files</b><br>Automatic OCR Strategy<br>Image Entity Extraction]
    B --> C[<b>Semantic Chunking</b><br>with Context Headers<br>Lines per Block: 100<br>Words per Chunk: 200<br>Model: gpt-4.1-2025-04-14]
    C --> D[<b>Embed Chunks</b><br>Model: text-embedding-3-large<br>Vector Size: 3072]
    D --> E[<b>Store Chunks</b><br>Local Qdrant database]
  end

  %% Query Pipeline
  subgraph Query
    F[<b>Ask Question</b>] --> G[<b>Embed Question</b><br>Model: text-embedding-3-large]
    G --> H[<b>Retrieve Nearest Chunks</b><br>Score Min: 0.1<br>Chunks Max: 40]
    E --> H
    H --> I[<b>Rerank by Relevance</b><br>Chunks Max: 30<br>Model: gpt-4.1-2025-04-14]
    I --> J[<b>Expand Context</b><br>Chunks Radius: 2]
    J --> K[<b>Generate Answer</b><br>Model: gpt-4.1-2025-04-14<br>Temperature: 1.1]
    K --> L[<b>Get Answer</b><br>in CLI, GUI, MCP]
  end

• 👉 Install Archive Agent on Linux
• 👉 Run Archive Agent
• 👉 MCP Tools
• 👉 Update Archive Agent

• Archive Agent
  • Find what you need with natural language
  • Natively index your documents on-device
  • Your AI, Your Choice
  • Scalable Performance
  • Architecture
  • Just getting started?
  • Documentation
  • Supported OS
  • Install Archive Agent
    • Ubuntu / Linux Mint
  • AI provider setup
    • OpenAI provider setup
    • Ollama provider setup
    • LM Studio provider setup
  • Which files are processed
  • How files are processed
  • OCR strategies
  • How smart chunking works
  • How chunk references work
  • How chunks are retrieved
  • How chunks are reranked and expanded
  • How answers are generated
  • How files are selected for tracking
  • Run Archive Agent
  • Quickstart on the command line (CLI)
  • CLI command reference
    • See list of commands
    • Create or switch profile
    • Open current profile config in nano
    • Add included patterns
    • Add excluded patterns
    • Remove included / excluded patterns
    • List included / excluded patterns
    • Resolve patterns and track files
    • List tracked files
    • List changed files
    • Commit changed files to database
    • Combined track and commit
    • Search your files
    • Query your files
    • Launch Archive Agent GUI
    • Start MCP Server
  • MCP Tools
  • Update Archive Agent
    • Archive Agent settings
      • Profile configuration
    • Watchlist
    • AI cache
  • Qdrant database
  • Developer's guide
    • Important modules
    • Network and Retry Handling
    • Qdrant Path Renaming Tool
    • Code testing and analysis
  • Known issues
  • Licensed under GNU GPL v3.0
  • Collaborators welcome
  • Learn about Archive Agent

Archive Agent has been tested with these configurations:

• Ubuntu 24.04 (PC x64)
• Ubuntu 22.04 (PC x64)

If you've successfully installed and tested Archive Agent with a different setup, please let me know and I'll add it here!

Please install these requirements before proceeding:

• Docker (for running Qdrant server)
• Python >= 3.10 (core runtime) (usually already installed)

This installation method should work on any Linux distribution derived from Ubuntu (e.g. Linux Mint).

To install Archive Agent in the current directory of your choice, run this once:

git clone https://github.com/shredEngineer/Archive-Agent
cd Archive-Agent
chmod +x install.sh
./install.sh

The install.sh script will execute the following steps:

• Download and install uv (used for Python environment management)
• Install the custom Python environment
• Install the spaCy model for natural language processing (pre-chunking)
• Install pandoc (used for document parsing)
• Download and install the Qdrant docker image with persistent storage and auto-restart
• Install a global archive-agent command for the current user

Archive Agent is now installed!

👉 Please complete the AI provider setup next.
(Afterward, you'll be ready to Run Archive Agent!)

Archive Agent lets you choose between different AI providers:

• Remote APIs (higher performance and cost, less privacy):

  • OpenAI: Requires an OpenAI API key.

• Local APIs (lower performance and cost, best privacy):

  • Ollama: Requires Ollama running locally.
  • LM Studio: Requires LM Studio running locally.

💡 Good to know: You will be prompted to choose an AI provider at startup; see: Run Archive Agent.

📌 Note: You can customize the specific models used by the AI provider in the Archive Agent settings. However, you cannot change the AI provider of an existing profile, as the embeddings will be incompatible; to choose a different AI provider, create a new profile instead.

If the OpenAI provider is selected, Archive Agent requires the OpenAI API key.

To export your OpenAI API key, replace sk-... with your actual key and run this once:

echo "export OPENAI_API_KEY='sk-...'" >> ~/.bashrc && source ~/.bashrc

This will persist the export for the current user.

💡 Good to know: OpenAI won't use your data for training.

If the Ollama provider is selected, Archive Agent requires Ollama running at http://localhost:11434.

• How to install Ollama.

With the default Archive Agent Settings, these Ollama models are expected to be installed:

ollama pull llama3.1:8b             # for chunk/rerank/query
ollama pull llava:7b-v1.6           # for vision
ollama pull nomic-embed-text:v1.5   # for embed

💡 Good to know: Ollama also works without a GPU. At least 32 GiB RAM is recommended for smooth performance.

If the LM Studio provider is selected, Archive Agent requires LM Studio running at http://localhost:1234.

• How to install LM Studio.

With the default Archive Agent Settings, these LM Studio models are expected to be installed:

meta-llama-3.1-8b-instruct              # for chunk/rerank/query
llava-v1.5-7b                           # for vision
text-embedding-nomic-embed-text-v1.5    # for embed

💡 Good to know: LM Studio also works without a GPU. At least 32 GiB RAM is recommended for smooth performance.

Archive Agent currently supports these file types:

• Text:
  • Plaintext: .txt, .md, .markdown
  • Documents:
    • ASCII documents: .html, .htm (images not supported)
    • Binary documents: .odt, .docx (including images)
  • PDF documents: .pdf (including images; see OCR strategies)
• Images: .jpg, .jpeg, .png, .gif, .webp, .bmp

📌 Note: Images in HTML documents are currently not supported.

📌 Note: Legacy .doc files are currently not supported.

📌 Note: Unsupported files are tracked but not processed.

Ultimately, Archive Agent decodes everything to text like this:

• Plaintext files are decoded to UTF-8.
• Documents are converted to plaintext, images are extracted.
• PDF documents are decoded according to the OCR strategy.
• Images are decoded to text using AI vision.
  • Uses OCR, entity extraction, or both combined (default).
  • The vision model will reject unintelligible images.
  • Entity extraction extracts structured information from images.
  • Structured information is formatted as image description.

See Archive Agent settings: image_ocr, image_entity_extract

Archive Agent processes files with optimized performance:

• Surgical Synchronization:
  • PDF analyzing phase is serialized (due to PyMuPDF threading limitations).
  • All other phases (vision, chunking, embedding) run in parallel for maximum performance.
• Vision operations are parallelized across images and pages within and across files.
• Embedding operations are parallelized across text chunks and files.
• Smart chunking uses sequential processing due to carry mechanism dependencies.

See Archive Agent settings: max_workers_ingest, max_workers_vision, max_workers_embed

For PDF documents, there are different OCR strategies supported by Archive Agent:

• strict OCR strategy (recommended):

  • PDF OCR text layer is ignored.
  • PDF pages are treated as images and processed with OCR only.
  • Expensive and slow, but more accurate.

• relaxed OCR strategy:

  • PDF OCR text layer is extracted.
  • PDF foreground images are decoded with OCR, but background images are ignored.
  • Cheap and fast, but less accurate.

• auto OCR strategy:

  • Attempts to select the best OCR strategy for each page, based on the number of characters extracted from the PDF OCR text layer, if any.
  • Decides based on ocr_auto_threshold, the minimum number of characters for auto OCR strategy to resolve to relaxed instead of strict.
  • Trade-off between cost, speed, and accuracy.

⚠️ Warning: The auto OCR strategy is still experimental. PDF documents often contain small/scattered images related to page style/layout which cause overhead while contributing little information or even cluttering the result.

💡 Good to know: You will be prompted to choose an OCR strategy at startup (see Run Archive Agent).

Archive Agent processes decoded text like this:

• Decoded text is sanitized and split into sentences.
• Sentences are grouped into reasonably-sized blocks.
• Each block is split into smaller chunks using an AI model.
  • Block boundaries are handled gracefully (last chunk carries over).
• Each chunk is prefixed with a context header (improves search).
• Each chunk is turned into a vector using AI embeddings.
• Each vector is turned into a point with file metadata.
• Each point is stored in the Qdrant database.

See Archive Agent settings: chunk_lines_block, chunk_words_target

💡 Good to know: This smart chunking improves the accuracy and effectiveness of the retrieval.

To ensure that every chunk can be traced back to its origin, Archive Agent maps the text contents of each chunk to the corresponding line numbers or page numbers of the source file.

• Line-based files (e.g., .txt) use the range of line numbers as reference.
• Page-based files (e.g., .pdf) use the range of page numbers as reference.

📌 Note: References are only approximate due to paragraph/sentence splitting/joining in the chunking process.

Archive Agent retrieves chunks related to your question like this:

• The question is turned into a vector using AI embeddings.
• Points with similar vectors are retrieved from the Qdrant database.
• Only chunks of points with sufficient score are kept.

See Archive Agent settings: retrieve_score_min, retrieve_chunks_max

Archive Agent filters the retrieved chunks .

• Retrieved chunks are reranked by relevance to your question.
• Only the top relevant chunks are kept (the other chunks are discarded).
• Each selected chunk is expanded to get a larger context from the relevant documents.

See Archive Agent settings: rerank_chunks_max, expand_chunks_radius

Archive Agent answers your question using the reranked and expanded chunks like this:

• The LLM receives the chunks as context to the question.
• LLM's answer is returned as structured output and formatted.

💡 Good to know: Archive Agent uses an answer template that aims to be universally helpful.

Archive Agent uses patterns to select your files:

• Patterns can be actual file paths.

• Patterns can be paths containing wildcards that resolve to actual file paths.

• 💡 Patterns must be specified as (or resolve to) absolute paths, e.g. /home/user/Documents/*.txt (or ~/Documents/*.txt).

• 💡 Use the wildcard * to match any file in the given directory.

• 💡 Use the wildcard ** to match any files and zero or more directories, subdirectories, and symbolic links to directories.

There are included patterns and excluded patterns:

• The set of resolved excluded files is removed from the set of resolved included files.
• Only the remaining set of files (included but not excluded) is tracked by Archive Agent.
• Hidden files are always ignored!

This approach gives you the best control over the specific files or file types to track.

💡 Good to know: At startup, you will be prompted to choose the following:

• Profile name
• AI provider (see AI Provider Setup)
• OCR strategy (see OCR strategies)

Screenshot of command-line interface (CLI):

For example, to track your documents and images, run this:

archive-agent include "~/Documents/**" "~/Images/**"
archive-agent update

To start the GUI, run this:

archive-agent 

Or, to ask questions from the command line:

archive-agent query "Which files mention donuts?"

To see the list of supported commands, run this:

archive-agent

To switch to a new or existing profile, run this:

archive-agent switch "My Other Profile"

📌 Note: Always use quotes for the profile name argument, or skip it to get an interactive prompt.

💡 Good to know: Profiles are useful to manage independent Qdrant collections (see Qdrant database) and Archive Agent settings.

To open the current profile's config (JSON) in the nano editor, run this:

archive-agent config

See Archive Agent settings for details.

To add one or more included patterns, run this:

archive-agent include "~/Documents/*.txt"

📌 Note: Always use quotes for the pattern argument (to prevent your shell's wildcard expansion), or skip it to get an interactive prompt.

To add one or more excluded patterns, run this:

archive-agent exclude "~/Documents/*.txt"

📌 Note: Always use quotes for the pattern argument (to prevent your shell's wildcard expansion), or skip it to get an interactive prompt.

To remove one or more previously included / excluded patterns, run this:

archive-agent remove "~/Documents/*.txt"

📌 Note: Always use quotes for the pattern argument (to prevent your shell's wildcard expansion), or skip it to get an interactive prompt.

To see the list of included / excluded patterns, run this:

archive-agent patterns

To resolve all patterns and track changes to your files, run this:

archive-agent track

To see the list of tracked files, run this:

archive-agent list

📌 Note: Don't forget to track your files first.

To see the list of changed files, run this:

archive-agent diff

📌 Note: Don't forget to track your files first.

To sync changes to your files with the Qdrant database, run this:

archive-agent commit

To see additional information on chunking and embedding, pass the --verbose option.

To bypass the AI cache (vision, chunking, embedding) for this commit, pass the --nocache option.

💡 Good to know: Changes are triggered by:

• File added
• File removed
• File changed:
  • Different file size
  • Different modification date

📌 Note: Don't forget to track your files first.

To track and then commit in one go, run this:

archive-agent update

To see additional information on chunking and embedding, pass the --verbose option.

To bypass the AI cache (vision, chunking, embedding) for this commit, pass the --nocache option.

archive-agent search "Which files mention donuts?"

Lists files relevant to the question.

📌 Note: Always use quotes for the question argument, or skip it to get an interactive prompt.

To see additional information on embedding, retrieval, reranking and querying, pass the --verbose option.

To bypass the AI cache (embedding, reranking) for this search, pass the --nocache option.

archive-agent query "Which files mention donuts?"

Answers your question using RAG.

📌 Note: Always use quotes for the question argument, or skip it to get an interactive prompt.

To see additional information on embedding, retrieval, reranking and querying, pass the --verbose option.

To bypass the AI cache (embedding, reranking) for this query, pass the --nocache option.

To launch the Archive Agent GUI in your browser, run this:

archive-agent gui

📌 Note: Press CTRL+C in the console to close the GUI server.

To start the Archive Agent MCP server, run this:

archive-agent mcp

📌 Note: Press CTRL+C in the console to close the MCP server.

💡 Good to know: Use these MCP configurations to let your IDE or AI extension automate Archive Agent:

• .vscode/mcp.json for GitHub Copilot agent mode (VS Code):
• .roo/mcp.json for Roo Code (VS Code extension)

Archive Agent exposes these tools via MCP:

📌 Note: These commands are read-only, preventing the AI from changing your Qdrant database.

💡 Good to know: Just type #get_answer_rag (e.g.) in your IDE or AI extension to call the tool directly.

This step is not immediately needed if you just installed Archive Agent. However, to get the latest features, you should update your installation regularly.

To update your Archive Agent installation, run this in the installation directory:

./update.sh

📌 Note: If updating doesn't work, try removing the installation directory and then Install Archive Agent again. Your config and data are safely stored in another place; see Archive Agent settings and Qdrant database for details.

💡 Good to know: To also update the Qdrant docker image, run this:

sudo ./manage-qdrant.sh update

Archive Agent settings are organized as profile folders in ~/.archive-agent-settings/.

E.g., the default profile is located in ~/.archive-agent-settings/default/.

The currently used profile is stored in ~/.archive-agent-settings/profile.json.

📌 Note: To delete a profile, simply delete the profile folder. This will not delete the Qdrant collection (see Qdrant database).

The profile configuration is contained in the profile folder as config.json.

💡 Good to know: Use the config CLI command to open the current profile's config (JSON) in the nano editor (see Open current profile config in nano).

💡 Good to know: Use the switch CLI command to switch to a new or existing profile (see Create or switch profile).

📌 Note: Since max_workers_vision and max_workers_embed requests are processed in parallel per file, and max_workers_ingest files are processed in parallel, the total number of requests multiplies quickly. Adjust according to your system resources and in alignment with your AI provider's rate limits.

The profile watchlist is contained in the profile folder as watchlist.json.

The watchlist is managed by these commands only:

• include / exclude / remove
• track / commit / update

Each profile folder also contains an ai_cache folder.

The AI cache ensures that, in a given profile:

• The same image is only OCR-ed once.
• The same text is only chunked once.
• The same text is only embedded once.
• The same combination of chunks is only reranked once.

This way, Archive Agent can quickly resume where it left off if a commit was interrupted.

To bypass the AI cache for a single commit, pass the --nocache option to the commit or update command (see Commit changed files to database and Combined track and commit).

💡 Good to know: Queries are never cached, so you always get a fresh answer.

📌 Note: To clear the entire AI cache, simply delete the profile's cache folder.

📌 Technical Note: Archive Agent keys the cache using a composite hash made from the text/image bytes, and of the AI model names for chunking, embedding, reranking, and vision. Cache keys are deterministic and change generated whenever you change the chunking, embedding or vision AI model names. Since cache entries are retained forever, switching back to a prior combination of AI model names will again access the "old" keys.

The Qdrant database is stored in ~/.archive-agent-qdrant-storage/.

📌 Note: This folder is created by the Qdrant Docker image running as root.

💡 Good to know: Visit your Qdrant dashboard to manage collections and snapshots.

Archive Agent was written from scratch for educational purposes (on either end of the software).

💡 Good to know: Tracking the test_data/ gets you started with some kind of test data.

To get started, check out these epic modules:

• Files are processed in archive_agent/data/FileData.py
• The app context is initialized in archive_agent/core/ContextManager.py
• The default config is defined in archive_agent/config/ConfigManager.py
• The CLI commands are defined in archive_agent/__main__.py
• The commit logic is implemented in archive_agent/core/CommitManager.py
• The CLI verbosity is handled in archive_agent/util/CliManager.py
• The GUI is implemented in archive_agent/core/GuiManager.py
• The AI API prompts for chunking, embedding, vision, and querying are defined in archive_agent/ai/AiManager.py
• The AI provider registry is located in archive_agent/ai_provider/ai_provider_registry.py

If you miss something or spot bad patterns, feel free to contribute and refactor!

Archive Agent implements comprehensive retry logic with exponential backoff to handle transient failures:

• AI Provider Operations: 10 retries with exponential backoff (max 60s delay) for network timeouts and API errors.
• Database Operations: 10 retries with exponential backoff (max 10s delay) for Qdrant connection issues .
• Schema Validation: Additional 10 linear retry attempts for AI response parsing failures with cache invalidation.
• Dual-Layer Strategy: Network-level retries handle infrastructure failures, while schema-level retries handle AI response quality issues.

This robust retry system ensures reliable operation even with unstable network conditions or intermittent service issues.

Archive Agent includes a dedicated tool for renaming file path prefixes in the Qdrant database:

./qdrant-rename-paths.py
# OR
uv run python qdrant-rename-paths.py

This interactive tool is useful when:

• You've moved your tracked files to a new directory structure
• You need to update file paths after reorganizing your documents
• You want to bulk-rename path prefixes without re-indexing all files

Features:

• Safe interactive workflow with preview and confirmation steps
• Connects to your currently active Archive Agent profile
• Shows matching files before making changes
• Updates all affected chunks in the Qdrant collection
• Preserves all embedding data and metadata

💡 Good to know: This tool modifies the Qdrant database directly, so ensure you have backups if working with critical data.

📌 Note:

• This tool will not update the tracked files. You need to update your watchlist (see Archive Agent settings) using manual search and replace.
• This tool will not rename the file://-prefixed names contained in the chunk context headers. Feel free to patch the script if needed.

To run unit tests, check types, and check style, run this:

./audit.sh

• While track initially reports a file as added, subsequent track calls report it as changed.

• Removing and restoring a tracked file in the tracking phase is currently not handled properly:

  • Removing a tracked file sets {size=0, mtime=0, diff=removed}.
  • Restoring a tracked file sets {size=X, mtime=Y, diff=added}.
  • Because size and mtime were cleared, we lost the information to detect a restored file.

• AI vision is employed on empty images as well, even though they could be easily detected locally and skipped.

• PDF vector images may not convert as expected, due to missing tests. (Using strict OCR strategy would certainly help in the meantime.)

• Binary document page numbers (e.g., .docx) are not supported yet; Microsoft Word document support is experimental.

• References are only approximate due to paragraph/sentence splitting/joining in the chunking process.

• AI cache does not handle AiResult schema migration yet. (If you encounter errors, passing the --nocache flag or deleting all AI cache folders would certainly help in the meantime.)

• Rejected images (e.g., due to OpenAI content filter policy violation) from PDF pages in strict OCR mode are currently left empty instead of resorting to text extracted from PDF OCR layer (if any).

• The spaCy model en_core_web_md used for sentence splitting is only suitable for English source text. Multilingual support is missing at the moment.

• HTML document images are not supported.

• The behaviour of handling unprocessable files is not customizable yet. Should the user be prompted? Should the entire file be rejected? Unprocessable images are currently tolerated and replaced by [Unprocessable image].

Copyright © 2025 Dr.-Ing. Paul Wilhelm <[email protected]>

This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

See LICENSE for details.

You are invited to contribute to this open source project! Feel free to file issues and submit pull requests anytime.

(Tap to open external link to YouTube playlist)