ezzuldinst/askme-bot

Purpose and Entry Points

AskMe is a Rust long-running bot for the Things social network. It detects @AskMe mentions, gathers thread/media context, asks Gemini for a response, and posts a threaded reply. The production entry point is src/main.rs; a developer dry-run path accepts --test-post <id> and can optionally publish with --post.

Runtime Orchestration

src/main.rs owns startup: loads environment credentials, initializes Things, Gemini, and optional Qdrant clients, then drives post handling. The bot requires Things and Gemini credentials; Qdrant initialization is non-fatal so processing continues without persistent memory. Domain structures passed across stages are defined in src/models.rs and src/entities.rs.

Things Network Boundary

src/things_client.rs encapsulates authentication to Things (email/password with OTP) and network operations needed to read posts, replies, and attached media, as well as publish the bot’s threaded response. This isolates the social-network API and its post/thread identifiers from orchestration and AI logic.

Context Construction and Conversation Semantics

The bot treats the mentioned post and its reply ancestry as conversational context. It limits included thread history through CONTEXT_DEPTH_LIMIT, enabling follow-up questions to retain prior discussion. src/entities.rs represents Things-side content/entities and supports output-oriented entity handling, while src/models.rs provides application-level message/request data structures.

Gemini Generation and Multimodal AI

src/gemini_client.rs is the boundary for Gemini chat and embedding APIs. It receives constructed conversation context plus available image/video material and returns the raw model response used for the Things reply. Gemini also supplies embeddings for semantic memory; its API key and model choices are configured by GEMINI_API_KEY, EMBEDDING_MODEL, and embedding dimension/batch settings.

Persistent Memory and Idempotency

src/qdrant_client.rs manages optional Qdrant-backed storage and retrieval. It supports semantic recall of prior thread content and persistent “already handled” state, preventing duplicate replies across restarts or network retries. src/qdrant_models.rs defines the Qdrant collection payload/vector representations. If Qdrant is unavailable, the bot degrades to stateless processing: no cross-restart deduplication and no recalled context.

Infrastructure and Operational Dependencies

The executable runs as a Cargo/Rust process (Cargo.toml) and calls external Things, Gemini, and Qdrant services. Qdrant is expected over local gRPC at port 6334 by default and can be started through scripts/start_qdrant.sh. Runtime secrets and endpoint/model tuning are supplied through environment variables, with example values in .env.example.