Hillock
An open-source deterministic retrieval tool that introduces the ability to judge "unanswerable queries" at the vector retrieval layer, enabling the system to explicitly refuse to answer rather than return spuriously relevant results. It is primarily used to improve the reliability of RAG systems.
Timeline (last 90 days)
Hillock is not positioned to replace vector databases, but to introduce deterministic judgment at the retrieval layer.
Hillock v1.0 was officially released after 6 months of testing.
The core goal of Hillock v1.0 is to let the system clearly identify and refuse unanswerable queries, rather than returning plausible-looking fragments.
Hillock adds a layer of judgment beyond similarity: whether the retrieval result truly constitutes effective support for the query.
Hillock's "determinism" does not mean completely eliminating probability, but rather making "unable to answer" a legitimate output.
Hillock is currently released as open source.
Hillock performs contextual semantic matching in a 10,000-dimensional hypervector space, belonging to the domain of hyperdimensional computing.
Hillock supports real-time model switching via the /model [name] command in the console without restarting the service.
Hillock is a local neuro-symbolic memory engine designed specifically for budget-constrained edge hardware, with a runtime memory footprint below 1.2 GB.
Hillock was shared by a developer in Reddit's Ollama community.
6 more timeline events
All Facts (15)
Hillock is currently released as open source.
50%UnverifiedHillock is not positioned to replace vector databases, but to introduce deterministic judgment at the retrieval layer.
50%UnverifiedHillock's "determinism" does not mean completely eliminating probability, but rather making "unable to answer" a legitimate output.
50%UnverifiedThe core goal of Hillock v1.0 is to let the system clearly identify and refuse unanswerable queries, rather than returning plausible-looking fragments.
50%UnverifiedHillock adds a layer of judgment beyond similarity: whether the retrieval result truly constitutes effective support for the query.
50%UnverifiedHillock uses Hebbian synaptic weights to establish association strengths between knowledge nodes, implementing an associative memory mechanism that requires no backpropagation training.
50%UnverifiedHillock performs contextual semantic matching in a 10,000-dimensional hypervector space, belonging to the domain of hyperdimensional computing.
50%UnverifiedHillock is a local neuro-symbolic memory engine designed specifically for budget-constrained edge hardware, with a runtime memory footprint below 1.2 GB.
50%UnverifiedHillock supports real-time model switching via the /model [name] command in the console without restarting the service.
50%UnverifiedHillock has been tested on a GTX 1070 GPU and a CPU-only laptop.
50%UnverifiedWhen a question cannot be answered by verified facts, Hillock's gating mechanism returns a hard-coded rejection response within 1 millisecond without invoking Ollama, thereby preventing large model hallucinations at the architectural level and saving GPU compute overhead.
50%UnverifiedHillock's TALON engine extracts structured knowledge triples locally without invoking any LLM, thereby avoiding GPU compute consumption during the text parsing phase.
50%UnverifiedHillock stores the extracted facts in a SQLite database.
50%UnverifiedHillock v1.0 was officially released after 6 months of testing.
50%UnverifiedHillock pushed version v0.6.0, bringing two capability upgrades: post-interaction MaxSim gating and multi-hop relational path reasoning.
50%