Module: PWN::MemoryIndex

Defined in:
lib/pwn/memory_index.rb

Overview

PWN::MemoryIndex is a lightweight local embedding index over PWN::Memory (~/.pwn/memory.json) so PromptBuilder can inject the N MOST-RELEVANT memories for the current request instead of the N newest. Embeddings prefer a direct Ollama /api/embed endpoint (PWN::Env[:ollama][:embed_model], default 'nomic-embed-text'). When only Open WebUI is configured, embeddings go through its /ollama/api/embed proxy with the openwebui key/base_uri.

Index layout (~/.pwn/memory.idx):

{ "<key>": { "sha": "<sha16 of value>", "vec": [Float,

Rebuilds are incremental: only entries whose value-sha changed are (re)embedded, so a warm index costs one embed call (the query).

Constant Summary collapse

INDEX_FILE = 
File.join(Dir.home, '.pwn', 'memory.idx')
DEFAULT_EMBED_MODEL = 
'nomic-embed-text'

Class Method Summary collapse

Class Method Details

.authors ⇒ Object

Author(s)

0day Inc. [email protected]



321
322
323
 
# File 'lib/pwn/memory_index.rb', line 321

public_class_method def self.authors
  "AUTHOR(S):\n  0day Inc. <[email protected]>\n"
end

.available? ⇒ Boolean

Supported Method Parameters

bool = PWN::MemoryIndex.available?

True when a local Ollama or Open WebUI base_uri is configured. All public methods degrade to substring recall when this is false.

Returns:



31
32
33
34
35
 
# File 'lib/pwn/memory_index.rb', line 31

public_class_method def self.available?
  !embed_endpoint.nil?
rescue StandardError
  false
end

.bm25(opts = {}) ⇒ Object 



275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
 
# File 'lib/pwn/memory_index.rb', line 275

public_class_method def self.bm25(opts = {})
  query = opts[:query].to_s.downcase
  docs = Array(opts[:docs]).map(&:to_s)
  qtok = expand_query(query: query)
  return docs.map { |d| { doc: d, score: 0.0 } } if qtok.empty?

  df = Hash.new(0)
  tokenized = docs.map do |d|
    toks = d.downcase.scan(/[a-z0-9]{3,}/)
    toks.uniq.each { |t| df[t] += 1 }
    toks
  end
  n = docs.length.to_f
  avgdl = tokenized.sum(&:length) / [n, 1].max
  k1 = 1.2
  b = 0.75
  ranked = docs.each_with_index.map do |doc, i|
    toks = tokenized[i]
    tf = Hash.new(0)
    toks.each { |t| tf[t] += 1 }
    dl = toks.length.to_f
    score = qtok.sum do |t|
      f = tf[t]
      next 0.0 if f.zero?

      idf = Math.log(((n - df[t] + 0.5) / (df[t] + 0.5)) + 1)
      idf * ((f * (k1 + 1)) / (f + (k1 * (1 - b + (b * dl / avgdl)))))
    end
    { doc: doc, score: score }
  end
  ranked.sort_by { |h| -h[:score] }
end

.embed(opts = {}) ⇒ Object

Supported Method Parameters

vecs = PWN::MemoryIndex.embed(texts: ['a', 'b'])

POST /api/embed on direct Ollama, or /ollama/api/embed via Open WebUI. Returns Array<Array> (one vector per input, nil on per-item failure). Batches of 32.



145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
 
# File 'lib/pwn/memory_index.rb', line 145

public_class_method def self.embed(opts = {})
  texts = Array(opts[:texts]).map(&:to_s)
  ep = embed_endpoint
  return Array.new(texts.length) unless ep && !texts.empty?

  model = ep[:model]
  browser = PWN::Plugins::TransparentBrowser.open(browser_type: :rest)
  rest    = browser[:browser]::Request
  headers = { content_type: 'application/json; charset=UTF-8' }
  token   = ep[:token]
  headers[:authorization] = "Bearer #{token}" if token && !token.to_s.empty?

  out = []
  texts.each_slice(32) do |batch|
    resp = rest.execute(
      method: :post,
      url: ep[:url],
      headers: headers,
      payload: { model: model, input: batch }.to_json,
      verify_ssl: false,
      timeout: 120
    )
    j = JSON.parse(resp, symbolize_names: true)
    out.concat(Array(j[:embeddings]))
  end
  out
rescue StandardError
  Array.new(texts.length)
end

.help ⇒ Object

Display Usage for this Module



327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
 
# File 'lib/pwn/memory_index.rb', line 327

public_class_method def self.help
  puts "USAGE:
    # Run available and return its result
    #{self}.available?

    # Run recall semantic and return its result
    #{self}.recall_semantic(
      query: 'required - user request / search text',
      limit: 'optional - top-K by cosine similarity (default 6)'
    )

    # Run to context and return its result
    #{self}.to_context(
      limit: 'optional - limit value consumed by #to_context (defaults to 6)',
      drop_hygiene_sops: 'optional - drop hygiene sops value consumed by #to_context',
      query: 'optional - search query string'
    )

    # Run refresh and return its result
    #{self}.refresh(
      mem: 'optional - mem value consumed by #refresh (defaults to PWN::Memory.load)'
    )

    # Run embed and return its result
    #{self}.embed(
      texts: 'required - Array texts value consumed by #embed'
    )

    # Rank documents by BM25 (pure Ruby; no embedder required).
    #{self}.bm25(
      query: 'required - search text',
      docs: 'required - Array of document strings'
    )

    # Run reset and return its result
    #{self}.reset

    # Print the AUTHOR(S) string for this module.
    #{self}.authors
  "
  constants.sort
end

.recall_semantic(opts = {}) ⇒ Object

Supported Method Parameters

hits = PWN::MemoryIndex.recall_semantic( query: 'required - user request / search text', limit: 'optional - top-K by cosine similarity (default 6)' )

Returns [{ key:, value:, category:, timestamp:, score: }, …] (newest-first Memory.recall shape + :score) or falls back to PWN::Memory.recall(query:) when embedding is unavailable.



47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
 
# File 'lib/pwn/memory_index.rb', line 47

public_class_method def self.recall_semantic(opts = {})
  query = opts[:query].to_s
  limit = (opts[:limit] || 6).to_i
  mem   = PWN::Memory.load
  return [] if mem.empty? || query.strip.empty?

  qv = embed(texts: [query]).first
  return fallback(query: query, limit: limit) unless qv

  idx = refresh(mem: mem)
  now = Time.now.utc
  # M2 — Generative-Agents retrieval: score = α·sim + β·recency + γ·importance
  # (Park '23). pwn had only sim; recency and importance now let a
  # 3-day-old human :fact outrank a 1-hour-old heuristic :lesson.
  scored = mem.map do |k, v|
    vec = idx.dig(k, :vec)
    next unless vec

    sim = cosine(a: qv, b: vec)
    age_h = (now - Time.parse(v[:timestamp].to_s)) / 3_600.0
    rec = Math.exp(-age_h / (24.0 * 7.0))
    imp = (v[:importance] || 0.5).to_f
    score = (0.6 * sim) + (0.25 * rec) + (0.15 * imp)
    { key: k, value: v[:value], category: v[:category], timestamp: v[:timestamp], score: score, sim: sim.round(3), importance: imp }
  rescue StandardError
    nil
  end
  scored.compact.sort_by { |h| -h[:score] }.first(limit)
rescue StandardError
  fallback(query: query, limit: limit)
end

.refresh(opts = {}) ⇒ Object

Supported Method Parameters

idx = PWN::MemoryIndex.refresh(mem: 'optional - preloaded PWN::Memory hash')

Incrementally (re)embed changed entries and prune deleted keys. Returns the in-memory index Hash keyed by memory key (Symbol).



118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
 
# File 'lib/pwn/memory_index.rb', line 118

public_class_method def self.refresh(opts = {})
  mem = opts[:mem] || PWN::Memory.load
  idx = load_index
  idx.delete_if { |k, _| !mem.key?(k) }

  todo = mem.reject { |k, v| idx.dig(k, :sha) == sha(text: v[:value].to_s) }
  unless todo.empty?
    vecs = embed(texts: todo.values.map { |v| "#{v[:category]}: #{v[:value]}" })
    todo.keys.each_with_index do |k, i|
      next unless vecs[i]

      idx[k] = { sha: sha(text: mem[k][:value].to_s), vec: vecs[i] }
    end
    save_index(idx: idx)
  end
  idx
rescue StandardError
  idx || {}
end

.reset ⇒ Object

Supported Method Parameters

PWN::MemoryIndex.reset



178
179
180
181
 
# File 'lib/pwn/memory_index.rb', line 178

public_class_method def self.reset
  FileUtils.rm_f(INDEX_FILE)
  {}
end

.to_context(opts = {}) ⇒ Object

Supported Method Parameters

ctx = PWN::MemoryIndex.to_context(query:, limit: 6)

Drop-in replacement for PWN::Memory.to_context that ranks by relevance to query instead of insertion order. Emitted format is identical so PromptBuilder needs no special-casing.



86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
 
# File 'lib/pwn/memory_index.rb', line 86

public_class_method def self.to_context(opts = {})
  limit = opts[:limit] || 6
  # Fetch extra when filtering hygiene SOPs so we still fill the budget.
  fetch_n = opts[:drop_hygiene_sops] ? [limit * 3, 18].max : limit
  hits = recall_semantic(query: opts[:query], limit: fetch_n)
  return PWN::Memory.to_context(limit: limit) if hits.empty?

  if opts[:drop_hygiene_sops]
    hits = hits.reject do |h|
      k = h[:key].to_s
      v = h[:value].to_s
      k.match?(/process_sop_.*(?:rubocop|rake|docs_after|code_hygiene)|operator_pref_docs_after_rubocop/i) ||
        v.match?(/\b(?:bundle exec )?rubocop\b.*\brake\b|docs_after_rubocop_rake|Documentation after code changes/i)
    end.first(limit)
    return PWN::Memory.to_context(limit: limit) if hits.empty?
  else
    hits = hits.first(limit)
  end

  ctx = "\n\nPERSISTENT MEMORY (relevance-ranked for this request - use PWN::Memory.remember to store new ones):\n"
  hits.each do |h|
    ctx += "- #{h[:key]} [#{h[:category]} @ #{h[:timestamp]}]: #{h[:value].to_s[0, 300]}\n"
  end
  ctx
end