TalkToDoc

By: Eden Oved

Introduction

TalkToDoc is a local AI-powered system that lets you "talk with your documents" - just like ChatGPT, but fully local and private.

It ingests PDF and Excel files, stores the text in a PostgreSQL database, builds a searchable TF-IDF index, and extracts structured project data into clean JSON files - including project timelines, milestones, contacts, keywords, and summaries.

The system minimizes LLM usage (under $3 total), caches all calls locally, and outputs both extracted answers and references to the source documents.

---

Pipeline Overview

1. Ingest all documents under data/ → stored in PostgreSQL + artifacts/pages.jsonl
2. Build Index using TF-IDF → artifacts/tfidf.pkl + DB page_vectors
3. Group documents by project ID
4. Pre-filter relevant pages per query
5. Extract structured fields using OpenAI (LLMClient)
6. Cache & track token cost → outputs/cost_log.jsonl
7. Output per-project structured summaries

---

Tools & Libraries

• Python 3.11
• PostgreSQL (via Docker)
• Typer (CLI)
• pandas
• PyMuPDF
• scikit-learn
• psycopg2
• openai
• Docker + docker-compose

---

Setup

1. Copy .env.example → .env and set:

OPENAI_API_KEY=sk-...
OPENAI_MODEL=gpt-4o-mini
TOKEN_BUDGET_DOLLARS=3.0
POSTGRES_HOST=db
POSTGRES_PORT=5432
POSTGRES_DB=talk_to_doc
POSTGRES_USER=postgres
POSTGRES_PASSWORD=postgres

2. Place your project folders and files under:

./data/

---

Run the System

docker compose up -d --build
docker compose exec app python app.py ingest
docker compose exec app python app.py build-index
docker compose exec app python app.py extract

Query Example:

You can ask questions in natural language, and the system will return relevant pages + evidence from your documents:

docker compose exec app python app.py query --q "project start date"

---

Reset (Clean Re-Run)

Removes all files under artifacts/ and outputs/ for a clean rerun.

docker compose exec app python app.py reset

---

Tests (optional)

Run automated pipeline tests (reset → ingest → index → extract → query).

docker compose exec app pytest -q

---

Output Files

File / Folder	Description
artifacts/pages.jsonl	Extracted text chunks
artifacts/tfidf.pkl	TF-IDF index + vectorizer
artifacts/cache/	Local cache of LLM responses
outputs/index.jsonl	Search index (debug/inspection)
outputs/manifest.jsonl	Summary of ingested documents
outputs/PRJ-*_key_params.json	LLM extraction project metadata
outputs/cost_log.jsonl	Token usage + LLM cost log

---

Built by Eden Oved
AI-driven document indexing and extraction project.