Project Intelligence Hub

Project Intelligence Hub turns scattered project materials into grounded answers, captures missing knowledge, and generates reusable project one-pagers.

Five Collaboration Stages

• connectors: stage 1 ownership for doc, ppt, markdown, csv, and text intake.
• rag_pipeline: stage 2 ownership for chunking, vector-store-ready persistence, retrieval, citations, and low-token answering.
• gap_identification: stage 3 ownership for confidence gaps, SME routing, and manual knowledge capture.
• enterprise_ui: stage 4 ownership for welcome, chat, gap capture, one-pager, and governance.
• evaluation_one_pager: stage 5 ownership for one-pager output and evaluation exports.

Implementation Support

• data: local inputs, processed chunks, captured knowledge, indexes, metrics, and validation logs.
• scripts: Windows-friendly commands for local development and demos.
• main.py: central FastAPI entrypoint and orchestrator for all five stages.
• platform_models.py: shared API/data contracts used by all five stages.
• platform_storage.py: shared local JSON storage, metrics, and validation logging.
• docs/COLLABORATION_GUIDE.md: single folder map, API surface, demo path, and collaboration rules.

Local Workflow

Copy-Item .env.example .env
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e .[dev]

Common commands:

make ingest
make api
make eval

On Windows without make, use the scripts in scripts/.

Run the POC:

.\.venv\Scripts\Activate.ps1
pip install -e .
.\scripts\run_api.ps1

Open http://127.0.0.1:8000, ingest data/samples/ncl-platform.txt, then ask:

What solution did Blend360 deliver for NCL?

Team Ownership

Area	Owner
Connectors	connectors
RAG pipeline	rag_pipeline
Gap identification	gap_identification
Enterprise UI	enterprise_ui
Evaluation and one-pager	evaluation_one_pager

Detailed member-by-member tasks, safe edit boundaries, and validation commands are in docs/COLLABORATION_GUIDE.md.

Core Demo Loop

1. Load project files.
2. Ask a question and show an answer with source citations.
3. Ask an unanswerable question and show a clear gap.
4. Capture a human-provided answer.
5. Ask again and show that the captured answer persists.
6. Generate one project one-pager from the available knowledge.
7. Export test-set answers as CSV or JSON.

Validation checks are shown in the UI and appended to data/processed/validation_log.jsonl. The runtime minimizes token usage by handling conversation locally and running retrieval before optional LLM synthesis.

Isolated Test Dataset

Place additional test files in data/raw/test, then run:

.\scripts\ingest_test.ps1

Test documents, chunks, and embeddings are written under data/processed/test and data/indexes/test. The primary corpus is left unchanged, while chat retrieval and one-pager generation search both datasets.

Optional LLM synthesis

ThinkAI works without an API key by returning grounded local evidence. To enable concise LLM synthesis, copy .env.example to .env and set OPENAI_API_KEY. OpenAI-compatible local services are also supported:

PIH_LLM_PROVIDER=ollama
PIH_LLM_MODEL=llama3.2:3b
PIH_LLM_BASE_URL=http://127.0.0.1:11434/v1