cd /news/ai-agents/show-hn-local-search-agent-offline-r… · home topics ai-agents article
[ARTICLE · art-57776] src=github.com ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

Show HN: Local Search Agent – offline RAG, no embeddings, free tier

Local Search Agent, a new Python framework, enables AI agents to search, fetch, and reason over local documents using BM25 keyword search via Meilisearch, avoiding cloud uploads, embeddings, and vector databases. The open-source tool offers a free tier with Google AI Studio or fully local setup via Ollama, addressing issues like stale indexes and black-box retrieval in traditional RAG systems.

read8 min views1 publishedJul 13, 2026
Show HN: Local Search Agent – offline RAG, no embeddings, free tier
Image: source

Give your AI agent a search engine for your local files.

Local Search Agent is a Python framework that gives your AI agent a search engine for your local files and lets it search, fetch, and reason over your local documents — the same way a researcher searches the web, but entirely on your machine.

Point it at a folder. Ask a question. The agent searches your documents, reads the relevant ones, and gives you an answer with citations — no cloud upload, no API calls to external search services, no embeddings, no vector stores.

"What was the AWS spend in Q3?"  →  agent searches index  →  fetches relevant docs  →  answers with sources

Traditional RAG (Retrieval-Augmented-Generation) has a fundamental problem: it converts your documents into embeddings and stores them in a vector database. That means:

Stale indexes— embeddings go out of date silently. You never know if the agent is reading your latest documents or a six-month-old snapshotBlack-box retrieval— you can't see why a document was retrieved or not. Debugging poor answers is guesswork** Chunking anxiety**— split too small and you lose context. Split too large and retrieval quality degrades. There's no right answer** Infrastructure overhead**— a vector database is another service to run, maintain, and pay for** Semantic drift**— embeddings are sensitive to how questions are phrased. A question about "cloud expenditure" may never match a document that says "AWS spend"

Local Search Agent takes a different approach: BM25 keyword search via Meilisearch, structured metadata, and a LangGraph agent loop with tools. The agent searches your document index the same way a developer searches Stack Overflow — with real queries, real results, and full transparency into what was retrieved and why.

The result is deterministic, auditable, and fast. You can see exactly what the agent fetched for every answer.

1. INGEST     Your documents → parsed, cleaned, chunked, indexed into Meilisearch
2. SERVE      FastAPI file server makes documents available to the agent via HTTP
3. SEARCH     LangGraph agent loop: search_local_index → fetch_local_url → reason
4. ANSWER     Agent returns an answer with inline source citations

Everything runs locally. Meilisearch downloads automatically on first use, no manual setup.

0.3.0 Release— Watch theRole-based Access Control** 0.2.1 Release**— Watch theZooming + Exporting conversation** 0.2.0 Release**— Watch theReranker + Watch mode** Native UI**— Watch theUI design and configuration video demo** CLI AGENT**— Watch theTerminal document querying video demo** Python API**— Watch theLocal Search Agent API Integration video demo

pip install local-search-agent
local-search config set-key --provider google --key YOUR_KEY

`ollama pull gemma4:e2b` (7.2GB) 
`ollama pull gemma4:e4b` (9.6GB) 
`ollama pull nemotron-3-nano:4b` (2.8GB Highly recommended)
local-search ui

The desktop UI open:

  • Create a workspace, name it, point it at a directory of files. The "Database path" field is optional — leave it blank to use the default location shown in the hint, or paste a custom path and click "Set & Restart".
  • Ingest (parse, clean, chunk).
  • Get a free google api key from ai-studio.
  • Set your api key at the top bar's right corner, or add a paid key for anthropic\openai . Note: For paid models or ollama, you will need to set model name via the config button at the top bar's right corner.
  • click Ingest from the left sidebar.
  • watch the progress bar at the bottom bar, wait until all files marked as completed.
  • Start asking questions.
local-search workspace create finance "C:\my_docs"
local-search ingest --workspace finance --dirs "C:\my_docs"

local-search serve --workspace finance

local-search query "What was the AWS spend in Q3?" --workspace finance --provider google

local-search query --workspace finance --provider google
python
from local_search_agent import SearchAgentFramework, SearchAgentConfig

config = SearchAgentConfig(
    document_dirs=["C:/my_docs"],
    workspace_name="finance",
    provider="google",
)

framework = SearchAgentFramework(config)
framework.ingest_and_index()
framework.start_file_server()

response = framework.query("What was the AWS spend in Q3?")
print(response["answer"])

Wrap an indexed workspace as a tool and plug it into any external AI agent — LangChain, LangGraph, Google Gemini SDK, or any framework that calls a function.

from local_search_agent import SearchAgentFramework, SearchAgentConfig, LocalSearchTool

config = SearchAgentConfig(
    document_dirs=["C:/skills"],
    workspace_name="skills",
    provider="google",
    model_name="gemini-3.1-flash-lite",  # cheap model for retrieval
)

framework = SearchAgentFramework(config)
framework.ingest_and_index()
framework.start_file_server()

skill_tool = LocalSearchTool(config)

from langchain_core.tools import tool

@tool
def search_skills(query: str) -> str:
    """Search the skills knowledge base for coding patterns and techniques."""
    return skill_tool.run(query).answer

Pass return_raw=True

to bypass the internal LLM summarisation and return the full document text verbatim — useful when the calling agent should reason over the raw content itself:

skill_tool = LocalSearchTool(config, return_raw=True)

See the Python API Reference for the full LocalSearchTool

documentation.

Running this for more than just yourself? Turn one shared deployment into a proper multi-user server with role-based access control — three roles (superadmin

/ admin

/ member

), admin

/member

granted per subject per workspace, superadmin

unconditional across the whole deployment, enforced on every protected route. Also includes optional per-role model/provider cost controls and concurrency/rate-limit management for cloud LLM accounts or shared Ollama hardware.

Three pluggable identity providers cover the common setups out of the box: a header provider for when you already have an authenticating reverse proxy in front, an API-key provider (with browser sessions) for when you don't have existing auth infrastructure, and a JWT provider that validates against your company's own IdP (Auth0, Okta, Azure AD, Google Workspace) when employees already sign in via SSO.

local-search workspace create finance "/srv/docs/finance"
local-search auth create-key --subject root@acme.com --display-name "IT/Ops" --superadmin
local-search auth create-key --subject alice@acme.com --display-name "Alice" --created-by root@acme.com
local-search grant-access --subject alice@acme.com --workspace finance --role admin

local-search ui --multi-tenant --db /var/lib/local-search-agent/prod.db

This is entirely opt-in — set identity_provider

on SearchAgentConfig

(or pass --multi-tenant

to local-search ui

) to turn it on; every existing single-user workflow above works exactly as written if you never do. See Role-Based Access Control for the full guide, and Production Deployment for running this as a shared server (Docker, systemd, reverse proxy) — note that Dockerfile/systemd/Caddy files live at the repo root, not in the PyPI package itself, so self-hosters should either clone this repo or copy the file contents straight out of that doc.

By default, scanned or image-based PDFs are processed using RapidOCR. Installing Tesseract enables a faster OCR path (~5 second per page vs. minutes without it).

Digitally-created PDFs (with a text layer) are never affected — they use direct text extraction and skip OCR entirely.

Windows

Download and run the installer from https://github.com/UB-Mannheim/tesseract/wiki

Make sure "Add Tesseract to the system PATH" is checked during installation.

Linux

sudo apt install tesseract-ocr        # Ubuntu / Debian
sudo dnf install tesseract             # Fedora / RHEL
sudo pacman -S tesseract               # Arch

macOS

brew install tesseract

After installation, restart the application — Tesseract is detected automatically. If it's not found, ingestion continues normally using RapidOCR with no errors.

Format Extension
.pdf
Word .docx
Excel .xlsx
PowerPoint .pptx
HTML .html , .htm
Plain text .txt , .md
CSV .csv
JSON .json
XML .xml
.eml

One command installpip install local-search-agent

. Meilisearch downloads automaticallyNo embeddings, no vector stores— BM25 search with structured metadata. Fast, deterministic, auditable** Native desktop UI**— pywebview window with live streaming agent responses, workspace management, and chat history** Multi-provider LLM**— Google, Ollama (local), OpenAI, Anthropic** Multi-workspace**— isolate document collections by department, project, channel, or topic. Each workspace is its own search index** Incremental sync**— background scheduler re-indexes only changed files. A 10,000-document corpus with 50 changes re-indexes only the 50** Multi-tenant RBAC**— opt-in three-role access control (superadmin

/admin

/member

) per workspace, with pluggable identity (header, API key, or JWT/SSO), plus optional per-role model/provider cost controls and concurrency/rate-limit management, for running one shared deployment across a teamFull CLI parity— everything you can do in the UI you can do from the terminal** Python API**— embed the framework directly in your own application** Cross-platform**— Windows, macOS, Linux

Guide Description

InstallationArchitectureCLI ReferencePython API ReferenceConfigurationIngestionMulti-WorkspaceRole-Based Access ControlProduction DeploymentSemantic SearchTroubleshootingContributions are welcome. Clone the repo and install in editable mode with dev dependencies:

git clone https://github.com/wiss84/local-search-agent.git
cd local-search-agent
pip install -e ".[dev]"

Run tests before submitting a PR:

pytest tests/ -v --cov=local_search_agent --cov-report=term-missing
ruff check .
ruff format .

MIT — see LICENSE for details.

Built by Wissam Metawee

── more in #ai-agents 4 stories · sorted by recency
── more on @local search agent 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/show-hn-local-search…] indexed:0 read:8min 2026-07-13 ·