Wires nomic-embed-text (iguana ollama) + pgvector on the shared
postgres18 into brain_query / brain_answer via Reciprocal Rank Fusion.
Pure BM25 stays the default; setting BRAIN_PG_DSN and BRAIN_EMBED_URL
together opts in. Setting one without the other is misconfiguration →
exit 1.
New packages:
- internal/embed
Client.Embed(ctx, text) → []float32 via POST {URL}/api/embed.
Defaults to nomic-embed-text:latest (768 dim). nil-on-empty-URL so
callers gate on a single nil check.
- internal/vectorstore
PGStore wraps a pgxpool against postgres18. Init creates
brain_embeddings(path PK, vector(768), updated_at) + HNSW cosine
index idempotently. Upsert / Delete / Search / KnownPaths.
Sync(brainDir, store, embedder) diffs brain/wiki/ against the store
and upserts new files / deletes removed ones; StartSync runs it on
a ticker (default 300s). Integration tests gated by BRAIN_PG_TEST_DSN.
- scripts/brain-embeddings-init.sql
One-time DBA setup: brain DB, brain_app role, vector extension,
GRANTs. Idempotent.
Search layer:
- search.QueryOptions gains Vector + Embedder fields.
- QueryContext is the cancellable variant; Query stays for callers.
- When both are set, BM25 (top-N) and pgvector (top-4N) candidates
merge via Reciprocal Rank Fusion (k=60, Cormack et al. 2009 — no
tuning knob, robust to scale differences between rankers).
- Vector-only hits are hydrated from disk so callers see uniform
Result records (path, title, excerpt, wing, hall, score).
- Wing/hall filters still apply to vector candidates via path-prefix.
- On embedder/vector errors the search falls back to BM25 — embedding
outage degrades quality but doesn't take the brain offline.
MCP wiring:
- mcp.Server.WithHybridRetrieval(v, e) opt-in setter, same shape as
WithReranker.
- brainQuery and brainAnswer pass the wired vector/embedder through
to search.QueryContext.
REST:
- POST /backfill-embeddings drives Sync synchronously. Returns
{added, deleted, errors[]}. 503 when feature is unconfigured.
cmd/server/main.go:
- BRAIN_PG_DSN + BRAIN_EMBED_URL together enable hybrid; one alone
→ exit 1.
- vectorAdapter bridges *PGStore (returns []Hit) to
search.VectorSearcher (which takes []VectorHit) without either
package importing the other.
- BRAIN_EMBED_SYNC_INTERVAL (default 300s) controls the background
Sync ticker.
Backend pivot from Qdrant to pgvector recorded in DECISIONS.md
2026-05-18 (supersedes 2026-04-08): postgres18 already runs in
databases/ ns, Qdrant was never deployed, one engine beats two.
Dependency: github.com/jackc/pgx/v5 — modern, native pgvector via
parametric vector literals.
Tests:
- embed.Client: empty-URL nil, request shape, dimension, upstream
error propagation, empty-text rejection.
- vectorstore.PGStore: dimension validation (unit); upsert/search/
KnownPaths (integration, BRAIN_PG_TEST_DSN-gated).
- vectorstore.Sync: adds new files, skips known, deletes
disappeared, skips _index.md, no-op when nil, collects embedder
errors.
- search.Query: hybrid promotes vector-only hits via RRF; falls
back to BM25 on embedder error.
Closes hyperguild#8.
158 lines
4.3 KiB
Go
158 lines
4.3 KiB
Go
package mcp
|
|
|
|
import (
|
|
"context"
|
|
"encoding/json"
|
|
"fmt"
|
|
"strings"
|
|
|
|
"github.com/mathiasbq/hyperguild/ingestion/internal/reranker"
|
|
"github.com/mathiasbq/hyperguild/ingestion/internal/search"
|
|
)
|
|
|
|
// rerankResults scores each candidate's excerpt against the query and
|
|
// returns up to top results whose score is positive, preserving the
|
|
// caller's input order (BM25 rank) within the kept set. The reranker is
|
|
// a filter: ties are broken by BM25, not by the reranker's binary score.
|
|
func rerankResults(ctx context.Context, rr *reranker.Client, query string, results []search.Result, top int) ([]search.Result, error) {
|
|
docs := make([]string, len(results))
|
|
for i, r := range results {
|
|
docs[i] = r.Excerpt
|
|
}
|
|
scores, err := rr.Score(ctx, query, docs)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
kept := make([]search.Result, 0, top)
|
|
for i, r := range results {
|
|
if scores[i] > 0 {
|
|
kept = append(kept, r)
|
|
}
|
|
if len(kept) == top {
|
|
break
|
|
}
|
|
}
|
|
return kept, nil
|
|
}
|
|
|
|
const (
|
|
answerSystemPrompt = `You are a knowledge assistant. Answer the question using ONLY the provided sources.
|
|
Cite source file paths inline when referencing specific content.
|
|
If the context does not contain enough information to answer, say so clearly.`
|
|
|
|
classifySystemPrompt = `Classify the document. Respond with JSON only, no markdown fences.
|
|
{"type":"...","title":"...","tags":["..."]}
|
|
Valid types: spec, plan, decision, note, wiki, log, code, unknown.`
|
|
)
|
|
|
|
type brainAnswerArgs struct {
|
|
Query string `json:"query"`
|
|
}
|
|
|
|
func (s *Server) brainAnswer(ctx context.Context, args json.RawMessage) (json.RawMessage, error) {
|
|
if s.answerLLM == nil {
|
|
return nil, fmt.Errorf("answer LLM not configured: set BRAIN_LLM_PRIMARY_URL")
|
|
}
|
|
var a brainAnswerArgs
|
|
if err := json.Unmarshal(args, &a); err != nil {
|
|
return nil, fmt.Errorf("parse args: %w", err)
|
|
}
|
|
if a.Query == "" {
|
|
return nil, fmt.Errorf("query is required")
|
|
}
|
|
|
|
// With reranker disabled: BM25 top-10 straight to the LLM.
|
|
// With reranker enabled: BM25 top-20 → cross-encoder filter → top-5.
|
|
bm25Limit := 10
|
|
if s.reranker != nil {
|
|
bm25Limit = 20
|
|
}
|
|
results, err := search.QueryContext(ctx, s.brainDir, search.QueryOptions{
|
|
Query: a.Query,
|
|
Limit: bm25Limit,
|
|
Vector: s.vector,
|
|
Embedder: s.embedder,
|
|
})
|
|
if err != nil {
|
|
return nil, fmt.Errorf("search: %w", err)
|
|
}
|
|
if s.reranker != nil && len(results) > 0 {
|
|
results, err = rerankResults(ctx, s.reranker, a.Query, results, 5)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("rerank: %w", err)
|
|
}
|
|
}
|
|
if len(results) == 0 {
|
|
return json.Marshal(map[string]any{
|
|
"answer": "No relevant content found in brain.",
|
|
"sources": []string{},
|
|
})
|
|
}
|
|
|
|
var sb strings.Builder
|
|
sources := make([]string, 0, len(results))
|
|
for _, r := range results {
|
|
fmt.Fprintf(&sb, "<source path=%q>\n%s\n</source>\n\n", r.Path, r.Excerpt)
|
|
sources = append(sources, r.Path)
|
|
}
|
|
|
|
answer, err := s.answerLLM(ctx, answerSystemPrompt, sb.String()+"Question: "+a.Query)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("llm: %w", err)
|
|
}
|
|
|
|
return json.Marshal(map[string]any{
|
|
"answer": answer,
|
|
"sources": sources,
|
|
})
|
|
}
|
|
|
|
type brainClassifyArgs struct {
|
|
Text string `json:"text"`
|
|
}
|
|
|
|
type classifyResult struct {
|
|
Type string `json:"type"`
|
|
Title string `json:"title"`
|
|
Tags []string `json:"tags"`
|
|
}
|
|
|
|
func (s *Server) brainClassify(ctx context.Context, args json.RawMessage) (json.RawMessage, error) {
|
|
if s.answerLLM == nil {
|
|
return nil, fmt.Errorf("answer LLM not configured: set BRAIN_LLM_PRIMARY_URL")
|
|
}
|
|
var a brainClassifyArgs
|
|
if err := json.Unmarshal(args, &a); err != nil {
|
|
return nil, fmt.Errorf("parse args: %w", err)
|
|
}
|
|
if a.Text == "" {
|
|
return nil, fmt.Errorf("text is required")
|
|
}
|
|
|
|
text := a.Text
|
|
if len(text) > 3000 {
|
|
text = text[:3000]
|
|
}
|
|
|
|
raw, err := s.answerLLM(ctx, classifySystemPrompt, text)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("llm: %w", err)
|
|
}
|
|
|
|
// Strip markdown fences if model adds them despite the instruction.
|
|
raw = strings.TrimSpace(raw)
|
|
raw = strings.TrimPrefix(raw, "```json")
|
|
raw = strings.TrimPrefix(raw, "```")
|
|
raw = strings.TrimSuffix(raw, "```")
|
|
raw = strings.TrimSpace(raw)
|
|
|
|
var cr classifyResult
|
|
if err := json.Unmarshal([]byte(raw), &cr); err != nil {
|
|
return nil, fmt.Errorf("parse classify response %q: %w", raw, err)
|
|
}
|
|
if cr.Tags == nil {
|
|
cr.Tags = []string{}
|
|
}
|
|
return json.Marshal(cr)
|
|
}
|