added a self-hosted metasearch engine to my PC this morning that runs in isolated containers, w/ a CPU-run reranker (ettin-68m) + privacy mode w/ Tor-only egress & a VPN tunnel over wireguard for torrent traffic. eat shit Google.
October 31, 2026 at 8:01 PM
added a self-hosted metasearch engine to my PC this morning that runs in isolated containers, w/ a CPU-run reranker (ettin-68m) + privacy mode w/ Tor-only egress & a VPN tunnel over wireguard for torrent traffic. eat shit Google.
A reranker scores search results more deeply to choose the best matches: say, the best five passages from a shortlist of 50. Before testing one, decide how much extra waiting would make better results no longer worth adopting. 1/5
What Is a Reranker? The Second-Stage Model That Improves Search Results
A reranker is a second-stage model that evaluates a small candidate set more deeply and reorders it for relevance after a fast initial retrieval. This guide explains the mechanism, trade-offs, evaluation, and controls th...
www.unite.ai
November 2, 2026 at 1:17 AM
A reranker scores search results more deeply to choose the best matches: say, the best five passages from a shortlist of 50. Before testing one, decide how much extra waiting would make better results no longer worth adopting. 1/5
+1. Reciprocal rank fusion is a cheap way to merge the two lists without tuning score scales, and it leaves the reranker less to do.
November 6, 2026 at 3:09 AM
+1. Reciprocal rank fusion is a cheap way to merge the two lists without tuning score scales, and it leaves the reranker less to do.
5️⃣ Cross-Encoder Reranking (Cohere Rerank / BGE-Reranker)
→ Retrieve the top 25 candidate chunks fast using hybrid search, then pass those 25 chunks through a high-precision Cross-Encoder Reranker to score and select the true top 5 before calling the LLM.
6️⃣ Metadata Pre-Filtering
→ Retrieve the top 25 candidate chunks fast using hybrid search, then pass those 25 chunks through a high-precision Cross-Encoder Reranker to score and select the true top 5 before calling the LLM.
6️⃣ Metadata Pre-Filtering
November 1, 2026 at 1:15 AM
5️⃣ Cross-Encoder Reranking (Cohere Rerank / BGE-Reranker)
→ Retrieve the top 25 candidate chunks fast using hybrid search, then pass those 25 chunks through a high-precision Cross-Encoder Reranker to score and select the true top 5 before calling the LLM.
6️⃣ Metadata Pre-Filtering
→ Retrieve the top 25 candidate chunks fast using hybrid search, then pass those 25 chunks through a high-precision Cross-Encoder Reranker to score and select the true top 5 before calling the LLM.
6️⃣ Metadata Pre-Filtering
Cornell/ASAPP paper: standard RAG retrieval + reranker fails when the right passage misses top-k. DecisionRAG instead navigates document structure with a lightweight decision model, beating dense retrieval + 8B reranker by 17.4 Hit@1 on Natural Questions.
November 2, 2026 at 1:47 PM
Cornell/ASAPP paper: standard RAG retrieval + reranker fails when the right passage misses top-k. DecisionRAG instead navigates document structure with a lightweight decision model, beating dense retrieval + 8B reranker by 17.4 Hit@1 on Natural Questions.
Keyword search or semantic search? Use both.
In Anthropic's study, combining them and adding a reranker cut retrieval failures by 67%.
Part 10 is out: how AI search works, plus a 20-query test to measure any search setup (yours included) 👇
In Anthropic's study, combining them and adding a reranker cut retrieval failures by 67%.
Part 10 is out: how AI search works, plus a 20-query test to measure any search setup (yours included) 👇
November 5, 2026 at 8:30 AM
Keyword search or semantic search? Use both.
In Anthropic's study, combining them and adding a reranker cut retrieval failures by 67%.
Part 10 is out: how AI search works, plus a 20-query test to measure any search setup (yours included) 👇
In Anthropic's study, combining them and adding a reranker cut retrieval failures by 67%.
Part 10 is out: how AI search works, plus a 20-query test to measure any search setup (yours included) 👇
DecisionRank: On-Policy Distillation for Decision-Model Rerankers
Meta presents a reranker that scores all candidates in one forward pass, trained on an LLM teacher's feedback on its own rankings.
📝 arxiv.org/abs/2610.42128
Meta presents a reranker that scores all candidates in one forward pass, trained on an LLM teacher's feedback on its own rankings.
📝 arxiv.org/abs/2610.42128
DecisionRank: On-Policy Distillation for Decision-Model Rerankers
Decision models target structured tasks by directly predicting choices or scores, avoiding autoregressive decoding latency. However, transferring a generative LLM's ranking knowledge to a decision mod...
arxiv.org
November 2, 2026 at 5:25 AM
DecisionRank: On-Policy Distillation for Decision-Model Rerankers
Meta presents a reranker that scores all candidates in one forward pass, trained on an LLM teacher's feedback on its own rankings.
📝 arxiv.org/abs/2610.42128
Meta presents a reranker that scores all candidates in one forward pass, trained on an LLM teacher's feedback on its own rankings.
📝 arxiv.org/abs/2610.42128
Our reranker comparison has no vendor speed ranking. The runs used different machines, dates and concurrency. A latency leaderboard would imply a controlled test we did not run. https://hevmind.com/writing/perplexity-decider-v1-2/?utm_source=bluesky&utm_campaign=perplexity-decider-v1-2
November 3, 2026 at 9:23 AM
Our reranker comparison has no vendor speed ranking. The runs used different machines, dates and concurrency. A latency leaderboard would imply a controlled test we did not run. https://hevmind.com/writing/perplexity-decider-v1-2/?utm_source=bluesky&utm_campaign=perplexity-decider-v1-2
A CPU reranker fits this self-hosted metasearch stack neatly. Container isolation and explicit Tor/WireGuard egress make the boundaries easy to inspect.
added a self-hosted metasearch engine to my PC this morning that runs in isolated containers, w/ a CPU-run reranker (ettin-68m) + privacy mode w/ Tor-only egress & a VPN tunnel over wireguard for torrent traffic. eat shit Google.
November 1, 2026 at 3:24 AM
A CPU reranker fits this self-hosted metasearch stack neatly. Container isolation and explicit Tor/WireGuard egress make the boundaries easy to inspect.
This Primo example makes the two-stage constraint concrete: the first pass must keep relevant papers in play so the reranker can lift them into the top results.
7/ Point here is Primo RA reranking of the top 30 makes the diff. The initial first stage retriever doesnt need to be perfect, just broad enough it does not drop relevant papers, reranker then gets the best results on top. This beats nicer looking boolean (eg ask anything) but poorer ranking!
October 30, 2026 at 11:22 PM
This Primo example makes the two-stage constraint concrete: the first pass must keep relevant papers in play so the reranker can lift them into the top results.
I swore I wasn’t going to go down this rabbit hole. Anyway, I got the metasearch engine running as a service on my NAS. Getting the local reranker is going to be the tricky part mostly because I need to figure out how to run the service, setup the model and get it talking to the search engine
added a self-hosted metasearch engine to my PC this morning that runs in isolated containers, w/ a CPU-run reranker (ettin-68m) + privacy mode w/ Tor-only egress & a VPN tunnel over wireguard for torrent traffic. eat shit Google.
October 31, 2026 at 11:55 PM
I swore I wasn’t going to go down this rabbit hole. Anyway, I got the metasearch engine running as a service on my NAS. Getting the local reranker is going to be the tricky part mostly because I need to figure out how to run the service, setup the model and get it talking to the search engine
Using ML to fix search could work if the training data was the old, indexable web rather than today's SEO sludge. RAG over historical snapshots (Wayback / Common Crawl) might recover the signal. Do you think an ML reranker trained on pre-AI-era results would beat current engines?
November 3, 2026 at 2:44 AM
Using ML to fix search could work if the training data was the old, indexable web rather than today's SEO sludge. RAG over historical snapshots (Wayback / Common Crawl) might recover the signal. Do you think an ML reranker trained on pre-AI-era results would beat current engines?
Retriever-free RAG from a Cornell/ASAPP paper replaces vector search with a decision model navigating a document tree. On Natural Questions it ranked the answer first 65.4% of the time, beating embedding search + reranker (48%) and GPT-5 reranking (60.8%).
November 2, 2026 at 5:05 PM
Retriever-free RAG from a Cornell/ASAPP paper replaces vector search with a decision model navigating a document tree. On Natural Questions it ranked the answer first 65.4% of the time, beating embedding search + reranker (48%) and GPT-5 reranking (60.8%).
1/ Terminology can be very slippery. Take the term "best match". Systematic review librarians know it in the context of Pubmed best match feature. It's a relevance ranking algo that takes your boolean set, ranks by Bm25, takes top 500 and reranks using a tuned ML reranker (LambdaMART). Simple right?
October 30, 2026 at 5:55 AM
1/ Terminology can be very slippery. Take the term "best match". Systematic review librarians know it in the context of Pubmed best match feature. It's a relevance ranking algo that takes your boolean set, ranks by Bm25, takes top 500 and reranks using a tuned ML reranker (LambdaMART). Simple right?
New research shows an LLM can rank its own passages via shared KV cache and short probes, using only a tiny LoRA plus pointer head. On tested sets it beats a dedicated reranker for evidence selection and downstream answer F1, though at higher latency and memory cost.
October 31, 2026 at 1:49 PM
New research shows an LLM can rank its own passages via shared KV cache and short probes, using only a tiny LoRA plus pointer head. On tested sets it beats a dedicated reranker for evidence selection and downstream answer F1, though at higher latency and memory cost.
Reranking gets expensive fast, and it can't fix retrieval that never surfaced the right results. Learn to build a multi-stage retrieval funnel instead.
Think you need a better reranker? Check your retrieval first.
Reranking gets expensive fast, and it can't fix retrieval that never surfaced the right results. Learn to build a multi-stage retrieval funnel instead.
bit.ly
November 2, 2026 at 4:00 PM
Reranking gets expensive fast, and it can't fix retrieval that never surfaced the right results. Learn to build a multi-stage retrieval funnel instead.
DecisionRAG skips the usual first-stage retriever and scores tree-structured corpus branches jointly in one forward pass. It lifts R@1 by 13.8 on MultiDoc2Dial and Hit@1 by 17.4 on Natural Questions over dense retrieval plus an 8B reranker, with no task-specific training and sub-2s latency on…
November 2, 2026 at 6:23 PM
DecisionRAG skips the usual first-stage retriever and scores tree-structured corpus branches jointly in one forward pass. It lifts R@1 by 13.8 on MultiDoc2Dial and Hit@1 by 17.4 on Natural Questions over dense retrieval plus an 8B reranker, with no task-specific training and sub-2s latency on…
Open a guide to rerankers (AI models that take a second, closer look at search results) and you'll find the same sentence repeated in all five of its numbered sections: the one explaining that a reranker isn't the first-stage retriever, the cheaper system that has to scan the whole collection 1/5
What Is a Reranker? The Second-Stage Model That Improves Search Results
A reranker is a second-stage model that evaluates a small candidate set more deeply and reorders it for relevance after a fast initial retrieval. This guide explains the mechanism, trade-offs, evaluation, and controls th...
www.unite.ai
October 29, 2026 at 1:11 PM
Open a guide to rerankers (AI models that take a second, closer look at search results) and you'll find the same sentence repeated in all five of its numbered sections: the one explaining that a reranker isn't the first-stage retriever, the cheaper system that has to scan the whole collection 1/5
Re3Rank: Listwise Reranking with Reasoning-Conditioned Retrieval Heads
Introduces a reranker that scores candidates via retrieval-head attention, using a generated reasoning trace instead of the question as the query.
📝 arxiv.org/abs/2611.02550
👨🏽💻 github.com/ryehr/RE3RANK
Introduces a reranker that scores candidates via retrieval-head attention, using a generated reasoning trace instead of the question as the query.
📝 arxiv.org/abs/2611.02550
👨🏽💻 github.com/ryehr/RE3RANK
Re$^3$Rank: Listwise Reranking with Reasoning-Conditioned Retrieval Heads
Attention-based rerankers score each candidate by the attention mass that retrieval heads route to it from the question, but the question, serving as a fixed query, carries limited information for loc...
arxiv.org
November 4, 2026 at 6:43 AM
Re3Rank: Listwise Reranking with Reasoning-Conditioned Retrieval Heads
Introduces a reranker that scores candidates via retrieval-head attention, using a generated reasoning trace instead of the question as the query.
📝 arxiv.org/abs/2611.02550
👨🏽💻 github.com/ryehr/RE3RANK
Introduces a reranker that scores candidates via retrieval-head attention, using a generated reasoning trace instead of the question as the query.
📝 arxiv.org/abs/2611.02550
👨🏽💻 github.com/ryehr/RE3RANK
0.19 s on CPU is basically free for a reranker, nice. Good to hear Brave and DDG hold up over Tor. Are you reranking the merged list from all engines, or each engine's results separately?
October 31, 2026 at 9:13 PM
0.19 s on CPU is basically free for a reranker, nice. Good to hear Brave and DDG hold up over Tor. Are you reranking the merged list from all engines, or each engine's results separately?
Nano / Small / Large have 0.27B / 1B / 4B activated parameters, built on T5Gemma2 and released under Apache 2.0.
The paper now calls the family Lychee-Reranker. All checkpoints:
huggingface.co/collections/...
The paper now calls the family Lychee-Reranker. All checkpoints:
huggingface.co/collections/...
Lychee-KaLM-Reranker&Jev - a KaLM-Embedding Collection
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
October 30, 2026 at 2:07 PM
Nano / Small / Large have 0.27B / 1B / 4B activated parameters, built on T5Gemma2 and released under Apache 2.0.
The paper now calls the family Lychee-Reranker. All checkpoints:
huggingface.co/collections/...
The paper now calls the family Lychee-Reranker. All checkpoints:
huggingface.co/collections/...
As far as I know AnythingLLM only does its built-in reranker (the "Accuracy Optimized" search setting on LanceDB), so a custom bge-reranker won't plug in without a proxy. Open WebUI lets you set your own reranking model under Documents with hybrid search on, if switching is an option.
October 30, 2026 at 8:11 AM
As far as I know AnythingLLM only does its built-in reranker (the "Accuracy Optimized" search setting on LanceDB), so a custom bge-reranker won't plug in without a proxy. Open WebUI lets you set your own reranking model under Documents with hybrid search on, if switching is an option.
Adilet Metinov, Chingiz Arziev: Glance: The Generator Is Its Own Reranker https://arxiv.org/abs/2610.40454 https://arxiv.org/pdf/2610.40454 https://arxiv.org/html/2610.40454
October 30, 2026 at 6:41 AM
Adilet Metinov, Chingiz Arziev: Glance: The Generator Is Its Own Reranker https://arxiv.org/abs/2610.40454 https://arxiv.org/pdf/2610.40454 https://arxiv.org/html/2610.40454
Or use SLERP to merge the same models:
It works for dense, sparse, reranker & multi-vector models via Arcee AI's mergekit.
pip install "sentence-transformers[merge]"
Compare the merge with both inputs on your data:
www.sbert.net/examples/sen...
It works for dense, sparse, reranker & multi-vector models via Arcee AI's mergekit.
pip install "sentence-transformers[merge]"
Compare the merge with both inputs on your data:
www.sbert.net/examples/sen...
October 29, 2026 at 2:36 PM
Or use SLERP to merge the same models:
It works for dense, sparse, reranker & multi-vector models via Arcee AI's mergekit.
pip install "sentence-transformers[merge]"
Compare the merge with both inputs on your data:
www.sbert.net/examples/sen...
It works for dense, sparse, reranker & multi-vector models via Arcee AI's mergekit.
pip install "sentence-transformers[merge]"
Compare the merge with both inputs on your data:
www.sbert.net/examples/sen...