<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>BM25 on OLAP Studio — Independent DuckDB Technical Blog</title><link>https://duckdblab.org/en/tags/bm25/</link><description>Recent content in BM25 on OLAP Studio — Independent DuckDB Technical Blog</description><generator>Hugo -- gohugo.io</generator><language>en-US</language><lastBuildDate>Thu, 19 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://duckdblab.org/en/tags/bm25/index.xml" rel="self" type="application/rss+xml"/><item><title>Full-Text Memory Recall Without a Vector DB: DuckDB FTS + BM25, Hands-On</title><link>https://duckdblab.org/en/post/duckdb-fts-memory-recall/</link><pubDate>Thu, 19 Nov 2026 00:00:00 +0000</pubDate><guid>https://duckdblab.org/en/post/duckdb-fts-memory-recall/</guid><description>&lt;img src="https://duckdblab.org/images/posts/duckdb-fts-memory-recall/cover.png" alt="Featured image of post Full-Text Memory Recall Without a Vector DB: DuckDB FTS + BM25, Hands-On" /&gt;&lt;p&gt;&lt;img alt="DuckDB FTS + BM25 memory recall: 10,000 memories indexed in 0.19s, one recall query in 0.035s" loading="lazy" sizes="(max-width: 767px) calc(100vw - 30px), (max-width: 1023px) 700px, (max-width: 1279px) 950px, 1232px" src="https://duckdblab.org/images/posts/duckdb-fts-memory-recall/cover.png"&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1: Measured DuckDB FTS extension results — a 10,000-entry memory table is indexed in 0.19 seconds, and a single BM25 recall query takes 0.035 seconds.&lt;/em&gt;&lt;/p&gt;
&lt;h2 id="the-more-memories-you-store-the-harder-they-are-to-retrieve"&gt;The more memories you store, the harder they are to retrieve
&lt;/h2&gt;&lt;p&gt;An agent&amp;rsquo;s memory pipeline has three stages: &lt;strong&gt;writing (what to keep), forgetting (what to drop), and recall (how to find it when you need it).&lt;/strong&gt; The first two decide what remains in the store; the last decides whether what remains is actually usable.&lt;/p&gt;
&lt;p&gt;Once memory accumulates to thousands or tens of thousands of entries, &amp;ldquo;retrieve the right few&amp;rdquo; becomes a real problem. The blunt instrument is &lt;code&gt;LIKE '%keyword%'&lt;/code&gt;: no relevance ordering (200 hits look the same as 2), synonyms and word-order variation all miss, and it degrades to a full scan as the table grows. The other extreme is a vector store for semantic recall — accurate, but it means running another service and paying another embedding bill, which is heavy equipment for a personal agent that just wants to work.&lt;/p&gt;
&lt;p&gt;The middle path is &lt;strong&gt;full-text search with BM25 ranking&lt;/strong&gt;. DuckDB ships an &lt;code&gt;fts&lt;/code&gt; extension (&lt;a class="link" href="https://duckdb.org/docs/2.0/core_extensions/full_text_search.html" target="_blank" rel="noopener"
 &gt;official docs&lt;/a&gt;): one PRAGMA builds the index, one &lt;code&gt;match_bm25&lt;/code&gt; call ranks the hits (BM25 is the &lt;a class="link" href="https://en.wikipedia.org/wiki/BM25" target="_blank" rel="noopener"
 &gt;classic Okapi retrieval model&lt;/a&gt;; DuckDB&amp;rsquo;s implementation follows the same spirit as SQLite FTS5). This post runs that path end to end with measured numbers: English recall, two real pitfalls (snapshot-only indexing, unsegmented Chinese text), and timings at the 10,000-row scale.&lt;/p&gt;

 &lt;blockquote&gt;
 &lt;p&gt;Environment: DuckDB v2.0.0 (Python, &lt;code&gt;INSTALL fts; LOAD fts;&lt;/code&gt;; the fts extension has been available since 1.5 with identical usage — numbers here are from 2.0.0). The sample memory table is synthetic (10 rows, modeled on real agent memory entries); every output below is an unedited measured result.&lt;/p&gt;

 &lt;/blockquote&gt;
&lt;h2 id="step-one-table-and-index"&gt;Step one: table and index
&lt;/h2&gt;&lt;div class="highlight"&gt;&lt;pre tabindex="0" class="chroma"&gt;&lt;code class="language-sql" data-lang="sql"&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="k"&gt;CREATE&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;TABLE&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;memories&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nb"&gt;INTEGER&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;created_at&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nb"&gt;DATE&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;kind&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nb"&gt;VARCHAR&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;content&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nb"&gt;VARCHAR&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="c1"&gt;-- insert 10 sample memories (an AdSense rejection, Payoneer payouts, blog analytics ...)
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="n"&gt;PRAGMA&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;create_fts_index&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;&amp;#39;memories&amp;#39;&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s1"&gt;&amp;#39;id&amp;#39;&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s1"&gt;&amp;#39;content&amp;#39;&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;overwrite&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;&lt;code&gt;create_fts_index&lt;/code&gt; creates a &lt;code&gt;fts_main_memories&lt;/code&gt; schema in the database; recall goes through its &lt;code&gt;match_bm25(document_id, query)&lt;/code&gt; macro: a hit returns a BM25 score, a miss returns NULL. So the recall query has a fixed shape — score in a subquery, filter and order outside:&lt;/p&gt;
&lt;div class="highlight"&gt;&lt;pre tabindex="0" class="chroma"&gt;&lt;code class="language-sql" data-lang="sql"&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="k"&gt;SELECT&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;score&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;content&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;FROM&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;SELECT&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;content&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;fts_main_memories&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;match_bm25&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s1"&gt;&amp;#39;adsense low value content&amp;#39;&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;AS&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;score&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;FROM&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;memories&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;WHERE&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;score&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;IS&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;NOT&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;NULL&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;ORDER&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;BY&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;score&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;DESC&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;Measured result for &lt;code&gt;adsense low value content&lt;/code&gt;:&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;id&lt;/th&gt;
 &lt;th&gt;score&lt;/th&gt;
 &lt;th&gt;content&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;1&lt;/td&gt;
 &lt;td&gt;2.7486&lt;/td&gt;
 &lt;td&gt;AdSense application for duckdblab.org was rejected for low value content&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;Of 10 rows, exactly one hits — and it is the right one. Two more queries behave the same way: &lt;code&gt;kdp royalties payment&lt;/code&gt; → id=2 (Payoneer payouts, 1.8324); &lt;code&gt;traffic revenue&lt;/code&gt; → id=3 (blog analytics, 1.6554). Even on a tiny corpus, BM25&amp;rsquo;s discriminative power is enough: irrelevant entries don&amp;rsquo;t even make the shortlist.&lt;/p&gt;
&lt;h2 id="pitfall-one-the-index-is-a-static-snapshot--new-memories-are-invisible"&gt;Pitfall one: the index is a static snapshot — new memories are invisible
&lt;/h2&gt;&lt;p&gt;The FTS index is a &lt;strong&gt;snapshot of the moment it was built&lt;/strong&gt;. Measured: insert a new memory after building the index —&lt;/p&gt;
&lt;div class="highlight"&gt;&lt;pre tabindex="0" class="chroma"&gt;&lt;code class="language-sql" data-lang="sql"&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="k"&gt;INSERT&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;INTO&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="n"&gt;memories&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="k"&gt;VALUES&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;11&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s1"&gt;&amp;#39;2026-11-10&amp;#39;&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s1"&gt;&amp;#39;fact&amp;#39;&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s1"&gt;&amp;#39;AdSense resubmission planned for November 16 after ads.txt authorization&amp;#39;&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;Query &lt;code&gt;adsense resubmission november&lt;/code&gt; immediately: the new row is &lt;strong&gt;absent&lt;/strong&gt;; only the old id=1 comes back (0.9162). Re-run &lt;code&gt;PRAGMA create_fts_index(..., overwrite=1)&lt;/code&gt; and query again:&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;id&lt;/th&gt;
 &lt;th&gt;score&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;11&lt;/td&gt;
 &lt;td&gt;2.6215&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;1&lt;/td&gt;
 &lt;td&gt;0.7180&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The new entry now ranks first. So an FTS-based memory recall path needs an explicit &amp;ldquo;rebuild index&amp;rdquo; step on the write path: rebuild after every write when the corpus is small (timings below make this cheap), and batched periodic rebuilds when it is large. If you expect it to maintain itself like a regular index, it will silently drop your newest memories — with no error, which is the hardest kind of bug to find.&lt;/p&gt;
&lt;h2 id="pitfall-two-unsegmented-chinese-text-gets-zero-hits"&gt;Pitfall two: unsegmented Chinese text gets zero hits
&lt;/h2&gt;&lt;p&gt;DuckDB FTS tokenizes on whitespace and punctuation. English is fine; an unsegmented Chinese sentence is one &amp;ldquo;token&amp;rdquo;, so:&lt;/p&gt;
&lt;div class="highlight"&gt;&lt;pre tabindex="0" class="chroma"&gt;&lt;code class="language-sql" data-lang="sql"&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="c1"&gt;-- notes table holding raw Chinese memory text, indexed:
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="c1"&gt;-- query &amp;#39;流量主 收入&amp;#39; → NO HITS
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="c1"&gt;-- query &amp;#39;流量主收入&amp;#39; → NO HITS
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;Both phrasings return nothing. The fix is to &lt;strong&gt;segment at ingest time&lt;/strong&gt;: split Chinese content into a space-separated token string, store it in a dedicated column (e.g. &lt;code&gt;content_seg&lt;/code&gt;), and index that column. Measured on the same data with the index pointed at the segmented column:&lt;/p&gt;
&lt;table&gt;
 &lt;thead&gt;
 &lt;tr&gt;
 &lt;th&gt;query&lt;/th&gt;
 &lt;th&gt;result&lt;/th&gt;
 &lt;/tr&gt;
 &lt;/thead&gt;
 &lt;tbody&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;code&gt;流量主 收入&lt;/code&gt;&lt;/td&gt;
 &lt;td&gt;id=1 (1.1107)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;tr&gt;
 &lt;td&gt;&lt;code&gt;公众号 质检&lt;/code&gt;&lt;/td&gt;
 &lt;td&gt;id=4 (0.8082) &amp;gt; id=1 (0.3197)&lt;/td&gt;
 &lt;/tr&gt;
 &lt;/tbody&gt;
&lt;/table&gt;
&lt;p&gt;The second row is the telling one: id=4 contains both &amp;ldquo;公众号&amp;rdquo; and &amp;ldquo;质检&amp;rdquo; and outranks id=1, which contains only &amp;ldquo;公众号&amp;rdquo; — BM25&amp;rsquo;s term-frequency and inverse-rarity weighting works on segmented Chinese exactly as it does in English. The segmenter can be anything (Jieba, a custom dictionary, even the LLM doing it on the write path); the point is it must happen at ingest, not at query time.&lt;/p&gt;
&lt;h2 id="scale-what-10000-memories-look-like"&gt;Scale: what 10,000 memories look like
&lt;/h2&gt;&lt;p&gt;We generated 10,000 synthetic rows (uniform word distribution, so we read timings, not scores):&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Build index: &lt;strong&gt;0.19 s&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;Single BM25 recall query: &lt;strong&gt;0.035 s&lt;/strong&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;These are numbers from one sandbox machine, not a universal constant, but the order of magnitude is the point: personal-agent memory scales (hundreds to tens of thousands of entries) are trivial for FTS. Rebuilding is cheap enough to do a full rebuild after every batched write — no incremental-index machinery required.&lt;/p&gt;
&lt;h2 id="when-not-to-use-it"&gt;When not to use it
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Semantic recall&lt;/strong&gt;: BM25 only sees literal terms. A memory that says &amp;ldquo;payout account&amp;rdquo; won&amp;rsquo;t match a query asking &amp;ldquo;how do I withdraw&amp;rdquo; — that is vector search&amp;rsquo;s job, and FTS cannot substitute. The pragmatic design is two-way recall: BM25 for exact terms, vectors for meaning, merge and dedupe. For the vector side we have hands-on &lt;a class="link" href="https://duckdblab.org/en/post/turbovec-offline-rag-local-first/" &gt;local RAG with turbovec&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Write-heavy workloads&lt;/strong&gt;: the index is a snapshot; high write rates mean high rebuild rates. Ten thousand rows at 0.19 s is nothing; a million rows is a different conversation. Check rebuild windows before committing.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Chinese without preprocessing&lt;/strong&gt;: as shown above, no segmentation, no index.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Cross-field weighting&lt;/strong&gt;: &lt;code&gt;create_fts_index&lt;/code&gt; indexes multiple columns, but per-field weight tuning is far less fine-grained than Elasticsearch/OpenSearch. Memory recall rarely needs that level; migrate only if you actually do.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="bottom-line"&gt;Bottom line
&lt;/h2&gt;&lt;p&gt;The memory pipeline is &lt;strong&gt;write (layering) → forget (decay) → recall (search)&lt;/strong&gt;. The first two answer &amp;ldquo;what stays in the store&amp;rdquo;; this one answers &amp;ldquo;can I find it when I need it&amp;rdquo;. DuckDB FTS sits in a clear place: zero new services, one PRAGMA to go live, millisecond-scale recall at the ten-thousand row mark; the costs are manual snapshot rebuilds, hand-rolled Chinese segmentation, and no semantic recall. For most personal agents, that trade is worth it.&lt;/p&gt;
&lt;h2 id="go-deeper"&gt;Go deeper
&lt;/h2&gt;&lt;ul&gt;
&lt;li&gt;&lt;a class="link" href="https://duckdblab.org/en/post/duckdb-ai-agent-memory-layer/" &gt;Building an AI Agent Memory Layer with DuckDB&lt;/a&gt; — the first post in this series: how memory is written, decayed, and merged (the vector-recall side); this one adds the literal-term recall (BM25) half&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://duckdblab.org/en/post/duckdb-ai-agent-brain/" &gt;Build an AI Agent with DuckDB as Its Brain&lt;/a&gt; — the memory table is one layer of agent state; this one covers letting an LLM query the store in natural language&lt;/li&gt;
&lt;li&gt;&lt;a class="link" href="https://duckdblab.org/en/post/turbovec-offline-rag-local-first/" &gt;turbovec Offline RAG: local embeddings + full-text two-way recall&lt;/a&gt; — the measured vector side of the &amp;ldquo;when not to use it&amp;rdquo; list above&lt;/li&gt;
&lt;/ul&gt;</description></item></channel></rss>