Skip to content

PRD.md

Latest commit

 

History

History
140 lines (114 loc) · 7.78 KB

PRD.md

File metadata and controls

140 lines (114 loc) · 7.78 KB
Raw

Personal LLM - Product Requirements Document

Status: v0.1 in active development Owner: Zaid Ali Syed Tagline: Your Second Brain. Your Digital Twin. Your Personal AI.

1. Executive Summary

Personal LLM is a local-first, privacy-preserving personal AI engine designed to know its user deeply, remember continuously, and eventually act autonomously on their behalf. Unlike ChatGPT, Claude, Gemini, or voice assistants (Siri/Alexa/Google Assistant), it is not a stateless chat product - it is infrastructure: a memory system, retrieval engine, model router, agent framework, and knowledge graph that this user's other projects (Recall, CivilizationOS, portfolio, future work) build on top of instead of re-implementing.

This document defines the full product vision (every subsystem from the original brief) and then triages it into what actually gets built now versus later, so ambition doesn't collapse into an unshippable everything-at-once project.

2. Vision

An AI that:

  • Knows the user - their projects, goals, preferences, writing/coding style.
  • Remembers - conversations, decisions, documents, and events accumulate into durable memory instead of vanishing at the end of a chat session.
  • Understands context and intent, not just literal commands.
  • Learns continuously from feedback and usage, without requiring model fine-tuning.
  • Protects the user's data - local-first, encryptable, auditable, under the user's control.
  • Works for the user - eventually completing multi-step tasks autonomously, with permission boundaries.

3. User Personas

Persona Description Primary needs
Zaid (primary) AI learner and builder, building public credibility via GitHub/LinkedIn, juggling multiple concurrent projects Remembers project context across sessions; surfaces forgotten tasks; answers questions grounded in his own notes/code, not generic web knowledge
Busy professional Manages email, calendar, docs across many tools Proactive summaries, deadline tracking, less context-switching
Researcher/learner Ingesting papers, courses, articles RAG over a growing personal knowledge base with citations, skill-gap tracking

4. Functional Requirements

Priority key: P0 = v0.1 (this build) · P1 = v0.2-v1.0 · P2 = post-v1 · P3 = speculative/research-gated, explicitly out of scope for now.

4.1 Memory

Requirement Priority
Ingest text/markdown/PDF into durable storage P0
Chunk + embed + semantic search (vector recall) P0
Episodic memory (timestamped events/interactions) P0
Importance scoring or memory items P0
Memory decay (old, rarely-accessed memories lose weight) P0
Memory consolidation / deduplication P0
Conflict resolution (contradicting facts) P1
Manual memory editing/deletion (user-facing) P1
Procedural memory ("how I like X done", reusable recipes) P1
Full-life ingestion (emails, calendar, files, messages, habits) P2

4.2 Reasoning

Requirement Priority
Grounded Q&A with citations from retrieved memory P0
Explicit uncertainty ("not in memory") vs hallucination avoidance P0
Self-critique / reflection pass on generated answers P1
Multi-step planning for agent tasks P1
Distinct "reasoning modes" (logical/scientific/creative/strategic) as separate models P3 - reject as stated. These are prompt strategies + verification passes on one strong model, not separate trained models. Building six bespoke reasoning engines is a research program, not a feature.

4.3 Model Orchestration

Requirement Priority
Single router interface for chat completion + embeddings P0
Cloud provider (Gemini free tier) P0
Optional local provider (Ollama) with automatic fallback P0
Confidence/verification signal on outputs P1
Multi-provider (paid models, OpenAI, etc. via router) P1
Automatic model selection by task complexity P2

4.4 Knowledge Graph

Requirement Priority
Entity/relation extraction from ingested content P0 (lite)
Node/edge storage, 1-hop traversal to enrich retrieval P0 (lite)
Rich graph across people/projects/files/calendar/emails P1-P2
Dedicated graph DB (Kùzu/Neo4j) at scale P2

4.5 Agents & Tools

Requirement Priority
Base agent abstraction (role + tools + memory scope + permissions) P1
Typed tool-calling layer with audit log P1
Specialized agents (research, coding, writing, email, etc.) P1-P2, added as config on the base agent, not new code
Fully autonomous multi-agent workforce managing the user's life P3 - real but distant; requires the permission/safety layer to mature first

4.6 Proactivity

Requirement Priority
Scheduled memory consolidation / weekly review job P1
Surfacing forgotten tasks, deadlines P1-P2
Full proactive life-management (spending anomalies, burnout detection, etc.) P2-P3, needs real integrations first

4.7 Voice & Vision

Requirement Priority
Text-only interaction (CLI, API, Streamlit chat) P0
Speech-to-text / text-to-speech P2
Real-time interruption, speaker recognition, voice cloning P3 - genuinely hard, high infra cost, low priority relative to memory/RAG value
Vision (screenshots, documents, whiteboards) P2
Live camera/desktop understanding P3

4.8 Integrations

Requirement Priority
Local filesystem ingestion P0
Gmail/Drive (via existing MCP connectors already available in this environment) P1
GitHub, Notion, Calendar, Slack, etc. P2
IoT/mobile P3

4.9 Developer Platform

Requirement Priority
Engine importable as a Python package by other local projects P0
Stable FastAPI surface for non-Python callers P1
Public SDK / plugin marketplace for third parties P3 - premature before there's a single trusted user (Zaid) fully served

5. Non-Functional Requirements

Category Requirement
Privacy All memory stored locally by default (SQLite + Chroma on-disk). No data leaves the device except explicit calls to a chosen cloud model provider.
Cost Must run end-to-end at $0 using free-tier Gemini + local embeddings. Paid providers opt-in only.
Offline capability Ingest, retrieve, and search must work with zero network access (local embeddings). Generation degrades gracefully (clear error) if no provider is reachable.
Latency Retrieval < 500ms for a personal-scale corpus (thousands of chunks). Generation latency bounded by provider.
Reliability Transient provider errors (429/5xx) auto-retried with backoff (pattern proven in resume-job-fit-ai/analyzer.py).
Portability All state in two portable artifacts: a SQLite file and a Chroma directory. Backup = copy data/.
Auditability Every memory write and (from v0.2) every tool action logged to an audit table.
Testability All LLM calls mockable; CI runs without secrets.

6. Success Metrics

Phase Metric
v0.1 Ingest a folder of real notes; ask a question; get a cited, grounded answer; correctly say "not in memory" for an absent fact; works with network disabled except for the generation call.
v0.2 An agent completes a 2+ step task using at least one tool, fully audited.
v1.0 Used daily by Zaid for at least one real task (e.g. querying his own project notes) for 2+ consecutive weeks.

7. Out of Scope (explicitly, for now)

Voice cloning, real-time multi-speaker voice, autonomous multi-agent "life management," public plugin marketplace, mobile/IoT integration, continuous model fine-tuning. These remain in the vision (see ROADMAP.md v2/v3) but are not designed in detail until v1 is real and used.