portfolio-agent

portfolio-agent is an open-source Python toolkit for turning a person's or team's documents, repositories, and web content into a citable, persona-aware assistant.

The repository is intentionally standardized around one supported path:

• ingest content with PortfolioAgent
• index it into the local FAISSVectorStore
• query it through the built-in persona-grounded RAGPipeline

What Is Supported

• Python SDK via PortfolioAgent
• Local FAISS indexing
• File ingestion for text/markdown/html/json/pdf
• GitHub and website ingestion through the SDK
• Querying with citations and lightweight session memory
• Optional FastAPI wrapper via create_app(...)

What Is Not the Supported Surface

• the old build_graph() entrypoint
• placeholder enterprise/platform claims
• legacy example scripts outside the portfolio demo
• unfinished HTTP endpoints for streaming, batch orchestration, admin, or security tooling

Public Release Scope

This repository is preparing for a narrow, truthful first public tag.

Release-ready expectations:

• install from source or package
• run the smoke/manual path
• run the smoke benchmark
• inspect source-backed answers and benchmark output

Not promised in this release:

• broad production guarantees
• exhaustive benchmark coverage for every ingestion format
• automatic validation of every optional runtime in every environment

Installation

When published as a package:

pip install portfolio-agent

From a fresh clone:

poetry install

From a locally built artifact:

poetry build
pip install dist/portfolio_agent-*.whl

For local embeddings, the package defaults to Hugging Face sentence-transformers. The first run may download the configured model.

Manual Verification

The fastest supported end-to-end check from a fresh clone is:

python scripts/manual_e2e.py --mode smoke

For the actual configured runtime path:

python scripts/manual_e2e.py --mode settings

See docs/MANUAL_E2E.md for the full SDK, CLI, and API verification flow. If your environment can load the real local embedding runtime, you can also run the opt-in integration check documented there.

Evaluation

The repo also includes a small benchmark for retrieval, grounding, abstention, and source-label quality on the canonical SDK path:

python scripts/run_evaluation.py --mode smoke

For the real configured runtime:

python scripts/run_evaluation.py --mode settings

See docs/EVALUATION.md for what the benchmark covers and what it does not.

Release Validation

For a maintainer/reviewer release pass:

python scripts/release_check.py

For a heavier clean-environment review of the configured embedding runtime:

python scripts/release_check.py --include-settings

See docs/RELEASE_CHECKLIST.md for the minimum release gate and review notes. The built-artifact dry run validates the wheel in a temporary virtual environment using available local site packages; it is meant to prove package usability, not to simulate a full online dependency-resolution install. For the end-to-end maintainer handoff, see docs/RELEASE_CANDIDATE.md.

Versioning

This project is currently using a pragmatic 0.x release shape:

• the canonical PortfolioAgent SDK path is the supported contract
• internals and heuristics may still evolve between minor releases
• release notes and changelog entries should be read as the source of truth for what changed

Current release candidate: 0.3.0rc1

See CHANGELOG.md for release notes.

Quick Start

from portfolio_agent import PortfolioAgent

agent = PortfolioAgent.from_settings()

agent.add_text(
    """
    Jane Doe is a backend engineer who works with Python, FastAPI, and retrieval systems.
    She has built developer tooling, API platforms, and AI product prototypes.
    """,
    source="profile.txt",
    document_type="txt",
)

result = agent.query("What does Jane work on?")
print(result.response)
print(result.sources)

Ingest a File

from portfolio_agent import PortfolioAgent

agent = PortfolioAgent.from_settings()
agent.add_file("sample_docs/portfolio.txt")

result = agent.query("Summarize the indexed portfolio")
print(result.response)

Ingest a GitHub Repo or Website

from portfolio_agent import PortfolioAgent

agent = PortfolioAgent.from_settings()
agent.add_github_repository("https://github.com/example/project")
agent.add_website("https://example.com")

FastAPI App

from portfolio_agent import PortfolioAgent, create_app

agent = PortfolioAgent.from_settings()
app = create_app(agent=agent)

Supported API endpoints:

• GET /api/v1/health
• POST /api/v1/query
• POST /api/v1/documents
• POST /api/v1/documents/file

CLI

The package exposes a portfolio-agent CLI:

portfolio-agent --add-file sample_docs/portfolio.txt --query "What are the key skills?"
portfolio-agent --interactive
portfolio-agent --serve

If you are validating a fresh clone and want the quickest repeatable check, use python scripts/manual_e2e.py --mode smoke before using the heavier local-embedding runtime.

Configuration

Key environment variables:

EMBEDDING_PROVIDER=hf
EMBEDDING_MODEL=sentence-transformers/all-MiniLM-L6-v2
EMBEDDING_BATCH_SIZE=16
# EMBEDDING_DEVICE=cpu
FAISS_INDEX_PATH=./faiss_index
REDACT_PII=true
TOP_K_RETRIEVAL=5

If you want OpenAI embeddings instead of local embeddings:

EMBEDDING_PROVIDER=openai
EMBEDDING_MODEL=text-embedding-3-small
OPENAI_API_KEY=your-key

Canonical Architecture

Source -> Ingestor -> Chunker -> Embedder -> FAISSVectorStore
                                     |
Query -> RouterAgent -> RetrieverAgent -> RerankerAgent -> PersonaAgent -> Response
                                                             |
                                                        MemoryManager

See docs/QUICKSTART.md for a slightly more detailed walkthrough.